Advancements in Non-Destructive Detection of Biochemical Traits in Plants Through Spectral Imaging-Based Algorithms: A Systematic Review

Abstract

This paper provides a comprehensive overview of the state of the art non-destructive methods for detecting plant biochemical traits through spectral imaging of leafy greens. It offers insights into the various detection techniques and their effectiveness. The review emphasizes the algorithms used for spectral data analysis, highlighting advancements in computational methods that have contributed to improving detection accuracy and efficiency. This systematic review, following the PRISMA 2020 guidelines, explores the applications of non-destructive measurements, techniques, and algorithms, including hyperspectral imaging and spectrometry for detecting a wide range of chemical compounds and elements in lettuce, basil, and spinach. This review focuses on studies published from 2019 onward, focusing on the detection of compounds such as chlorophyll, carotenoids, nitrogen, nitrate, and anthocyanin. Additional compounds such as phosphorus, vitamin C, magnesium, glucose, sugar, water content, calcium, soluble solid content, sulfur, and pH are also mentioned, although they were not the primary focus of this study. The techniques used are showcased and highlighted for each compound, and the accuracies achieved are presented to demonstrate effective detection.

1. Introduction

The increase in global population has contributed to the constantly growing demand for high-quality, nutrient-rich leafy greens such as lettuce, basil, and spinach [1]. This growing demand intensifies the need for agricultural practices that align with the United Nations sustainability goals [2], ensuring an adequate supply of high-quality, nutrient-rich leafy greens. This is also driven by consumers with growing awareness of health and sustainability [1]. Advances in agriculture, such as greenhouses, vertical farming, and smart farming, have presented new opportunities in optimizing crop quality and yield. A critical tool for optimization is monitoring and measurements of the chemical compounds in leafy greens, which can be an indicator of nutritional value, marketability, and shelf life. However, traditional methods for assessing chemical compounds in plants, such as destructive sampling and laboratory analysis, are time-consuming, labor-intensive, and often impractical for real-time or large-scale applications.

Spectral imaging techniques, such as hyperspectral imaging, multispectral imaging, and spectrometry, have emerged as powerful non-destructive tools for plant phenotyping [3,4]. By analyzing the interaction that occurs between light and plant tissue, these techniques can enable accurate and rapid detection of crucial chemical compounds in plants [3]. By capturing spectral information, those methods can quantify and map physiological parameters of the plants [5]. The non-invasive nature of this approach makes it suitable for monitoring plant health and quality [5,6].

The rationale for this systematic review is to summarize recent developments in spectral imaging technologies and their applications in detecting chemical compounds in leafy greens. Although many studies have explored this area individually, there is a need to bring together existing findings on the effectiveness and accuracy of spectral imaging techniques in agricultural settings. Therefore, this paper aims explicitly to address this gap by systematically reviewing recent studies on the non-destructive detection of chemical compounds in leafy greens using spectral imaging techniques.

This paper was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines to ensure a rigorous and transparent review process. By collecting the results of recent studies, it aims to provide a comprehensive database of the techniques and accuracies achieved in detecting chemical compounds in most common plants.

The time frame reflects the rapid advancements in sensor technologies, machine learning algorithms, and computation methods. Key chemical compounds of interest include chlorophyll, carotenoids, anthocyanins, nitrogen, and nitrate. In this review, there are also compounds that were not directly searched but additionally added from studies included, such as vitamin C, magnesium, sulfur, phosphorus, water content, calcium, cellulose, sugar, and soluble solid content.

This paper is structured as follows: Section 2 contains all information about data collection, search methodology for articles, and selection processes. In Section 3, technologies of hyperspectral cameras, multispectral cameras, and spectrometers are discussed. Section 4 is divided into five subsections: Section 4.1, Section 4.2, Section 4.3, Section 4.4 and Section 4.5. All subsections contain information about chemical compounds detected and tables containing data collected. Last, Section 5 talks about findings and potential future work.

2. Materials and Methods

2.1. Data Collection

This systematic review aims to provide a comprehensive analysis of studies focusing on the detection of chemical compounds and elements in lettuce, spinach, and basil using spectrometers and hyperspectral cameras. The review consolidates state-of-the-art research, presenting the achieved accuracies and the methodologies employed to attain them. Relevant studies were identified through a search of three databases, Scopus, Web of Science, and Google Scholar, yielding an initial collection of 527 studies. Following a detailed screening process, 28 studies were selected for inclusion. The analysis concentrated on publications from 2019 onward and provided a synthesis of key parameters, including the accuracies achieved, the methods applied, the spectral ranges utilized, the number of images acquired, data dimensionality reduction algorithms, and the plant species studied.

2.1.1. Search Methodology

The articles were retrieved from multiple databases, including Scopus, Google Scholar, and Web of Science. To ensure consistency and relevance in the search results, a set of base keywords was used: (“Hyperspectral” OR “Spectroscopy”) AND (“Lettuce” OR “Spinach” OR “Basil”) AND PUBYEAR > 2019 AND PUBYEAR < 2025. Additional keywords related to specific chemical compounds and elements of interest, such as chlorophyll, carotenoids, anthocyanin, nitrogen, nitrate, and nitrite, were sequentially incorporated into the search. For Scopus and Web of Science, the database results were exported and subsequently uploaded to the Rayyan [7] software for systematic screening and selection. In the case of Google Scholar, where search queries frequently yielded tens of thousands of results, a tailored screening methodology was implemented. The search process involved reviewing results page by page, identifying relevant articles by reading titles and abstracts, and determining a stopping criterion. Specifically, the search continued until n pages, where n represents the last page containing at least one relevant article, with an additional page (+1) screened to ensure no relevant studies were overlooked. This structured approach allowed for efficient handling of large volumes of search results while maintaining the integrity and relevance of the selected studies. Studies were deemed eligible for synthesis if they matched predefined search and inclusion criteria explicitly.

2.1.2. Selection Process

The search yielded 527 articles; only studies focusing specifically on lettuce, basil, or spinach were advanced to further stages of analysis. Of the 527 articles, 281 were deleted due to being duplicates. After applying these criteria, 246 papers were selected for an in-depth screening process. The in-depth screening process involved a thorough review of all titles and abstracts. For Google Scholar, this step was already completed during the initial search. To ensure relevance to sensing technologies, only studies utilizing spectral ranges between 200 nm and 2500 nm were included, while papers based solely on RGB imaging were excluded. Articles that did not detect or estimate biochemical traits were also excluded. For papers where the relevance could not be determined solely from the title and abstract, the full content of the paper was examined to determine whether it should be included or excluded. This systematic approach ensured that only the most relevant studies were included for detailed analysis. In the final exclusion, 60 papers were removed because they were review papers, not relevant, or not written in English. The PRISMA 2020 flow diagram of the screening proces is shown in Figure 1. The plant type, spectral wavelengths, analytical methods, achieved accuracies, and the number of samples utilized across all accepted studies were meticulously gathered and analyzed. This information was organized into tables, categorized by the specific chemical compounds or elements of interest. Tables were deliberately chosen for their clarity and comparative utility due to significant methodological heterogeneity. Heterogeneity among studies was qualitatively described through structured tabular comparison rather than quantitatively via meta-analysis. The table contents are described in the following subsection.

2.1.3. Table Contents

The tables were designed to provide a detailed and comprehensive overview, enabling clear and direct comparisons between studies in terms of the methodologies employed, the spectral ranges used for detection, the accuracy of the results obtained, and dimensionality reduction techniques if used. Furthermore, the tables not only summarize the key elements of each study but also identify variations in experimental designs. These include factors such as the number of samples included in each study, the types of imaging techniques employed, and any other notable methodological differences. For more information about the algorithms used, please refer to the Abbreviations. This organized presentation of the data offers valuable insights into the current state of the art approaches for the non-destructive detection of chemicals and elements in leafy greens. By systematically compiling and comparing this information, the tables serve as a resource for identifying the most reliable and innovative methods in the field, providing a clear picture of the progress and future directions in spectral imaging for plant biochemical detection.

2.1.4. Risk of Bias

A simplified Risk of Bias assessment was performed for each included study using a modified QUADAS-2 approach, focusing on clarity of methods description, reference biochemical measurements, validation techniques, and reproducibility of reported results. Each study was independently assessed by two reviewers, and discrepancies were resolved through discussion.

3. Sensor Technologies in Chemical and Elemental Detection

Detection of chemical compounds and elements in plants often relies on laboratory techniques. While these methods can yield high accuracy and provide quantitative information, they are inherently destructive. They require plant tissue to be sampled, processed, and often destroyed. These techniques limit the possibility of repeated measurements on the same plant and increase labor, time, and cost, making them less feasible for large-scale testing or real-time monitoring of plants [8]. In contrast, non-destructive sensing technologies, such as hyperspectral cameras and spectrometers, have been shown to be powerful alternatives. These tools allow for the detection of plant chemical composition without damaging the plant, enabling repeated measurements of the same plant over time and preserving the plant’s integrity. Hyperspectral imaging, in particular, captures a wide range of spatial and spectral data simultaneously [8]. Similarly, spectrometers provide precise spectral information at a specific point on the plant, enabling focused and detailed analyses. These non-destructive methods not only accelerate data acquisition but also support sustainable and efficient monitoring, making them invaluable in modern agricultural research and production systems. The next subsections explore the aforementioned technologies in more detail.

Spectral measurements indoors and outdoors both introduce challenges. While Indoor measurements are inherently more controlled, they still require a strict illumination setup, usually a combination of LED and fluorescent lighting, which can introduce noise. Outdoor measurements introduce variability in lighting conditions, such as angle of light and cloud cover, requiring an advanced real-time algorithm. It can also be affected by temperature, humidity, and atmospheric interference.

3.1. Hyperspectral Imaging

Hyperspectral cameras are high-tech image devices that are able to capture a wide range of the electromagnetic spectrum. Contrary to conventional RGB cameras that capture only three bands (red, blue, and green) and instantly combine them, hyperspectral camera captures a dense sampling of the spectrum range, usually hundreds of narrow bands, which provides rich spectral information about the target in each pixel. These systems can cover wavelengths from ultraviolet to thermal infrared and provide a unique insight into plant biochemical traits while recording a dense spectral signature for each pixel [9]. Apart from plant biochemical traits detection hyperspectral cameras have wide application in remote sensing where they can be used to asses water quality materials [9], in biochemical for non-invasive medical diagnostics [10], quality control in industry [11], cultural heritage preservation [12], military and security [13], and many more. But despite the potential, hyperspectral cameras face challenges such as high cost, and imaging generates large datasets, requiring advanced algorithms for efficient storage and analysis [12]. Future advancements are being focused on, making cheaper cameras with wider spectral range and miniaturization of the systems.

Data dimensionality reduction algorithms are usually needed for working with hyperspectral images due to the large volumes of data generated with every picture, which can become computationally heavy and hard to store. Techniques like PCA, PLSR, or CARS, among many, have proven to be crucial in reducing the complexity of generated data, reducing noise and computational load, which enables faster and more efficient processes while maintaining or increasing accuracy.

3.2. Multispectral Imaging

Multispectral cameras capture images with multiple wavelength bands; these images offer wider spectral information than RGB images. The cameras are cheaper and usually smaller than hyperspectral cameras, which helps with being more widely used in remote sensing [14]. More specifically, these cameras are widely used for environmental monitoring, where they identify vegetation health, soil conditions, and water quality with high precision by leveraging their ability to capture data across narrow spectral bands [14]. With a light weight and lower cost, they are a primary choice to be spaceborne and airborne. Multispectral camera technology is continuously advancing, enabling the development of smaller devices with a wider range of bands and more affordable prices.

3.3. Spectrometers

Spectrometers are a valuable tool for analyzing the properties of leaves. Several types of spectrometers are mostly used for close-proximity measurements. Fiber optic spectrometers are used due to their flexibility and adaptability. These devices employ fiber optic cables for light transmission from the leaf to the device. This design is particularly suitable for field studies because it ensures minimal disturbance of the sample, and it is possible to measure in hard-to-reach areas [15]. Spectrometers provide the absorption spectrum of the target, which shows how much light is absorbed at each wavelength. By measuring multiple points on the leaf, users can have a good overview of the leaf’s spectral data.

3.4. Data Acquisition Methods in Spectral Imaging: Principles and Instrumentation

Spectrometry, multispectral imaging, and hyperspectral imaging are distinct yet complementary spectral measurement techniques used for biochemical trait detection. Spectrometry typically involves point measurements, offering high spectral resolution and precision but lacking spatial context. In contrast, multispectral and hyperspectral imaging techniques capture spatially resolved spectral data. Multispectral imaging captures data in selected discrete bands, providing faster acquisition speed, lower cost, and simplicity, but at the expense of reduced spectral resolution. Hyperspectral imaging, on the other hand, acquires contiguous spectral data across many narrow bands, enabling detailed biochemical characterization but resulting in higher costs, increased complexity, and larger data volumes.

Table 1. Distinctions between spectrometry, multispectral, and hyperspectral imaging.

Technique	Data Acquisition Principle	Cost	Dimensions	Data Completeness	Acquisition Speed	Advantages	Disadvantages
Spectrometry [16]	Point measurement	Low	Small	Moderate	Fast	Low cost, high precision	No spatial information
Multispectral Imaging [17]	Snapshot/Filter-based	Medium	Compact	Limited	Very fast	Low cost, portable	Limited spectral resolution
Hyperspectral Imaging [18]	Pushbroom/Line scanning	High	Larger	High	Slow to moderate	High spectral and spatial resolution	Expensive, computationally intensive

below summarizes the key differences among these techniques, highlighting their acquisition principles, advantages, and limitations to guide practical decision making and methodological selection.

4. Results

In this section, various spectral imaging and spectrometry techniques are shown for their effectiveness in non-destructive biochemical detection, particularly in basil, spinach, and lettuce. These chemical compounds play essential roles in plant health, development, and nutritional quality, making their accurate and non-destructive quantification crucial for both growers and researchers. These findings underscore the potential of combining hyperspectral imaging and advanced regression models for precise and non-destructive biochemical quantification in plants. In the articles collected wide range of spectral technologies, including both hyperspectral imaging systems and spectrometers. Wavelengths captured by hyperspectral cameras ranged between 350 nm–2500 nm, the most used ones ranging between 400 nm to 1000 nm. For the spectrometers ranges used were 305 nm–2500 nm, with many using a range in NIRS.

The result section contains all the data collected in table format. This section is divided by biochemical trait. Every subsection gives a quick overview of the chemical compound combined with tables. Most included studies clearly described spectral imaging methodologies and validation approaches (low risk), but several studies had unclear descriptions of biochemical reference measurements (unclear or high risk). A summary of these assessments is presented in

Table 2. Risk of Bias.

Study	How Clearly the Spectral Imaging Methodology Was Described?	How Clearly the Reference Biochemical Methods (e.g., Standard Laboratory Measurements) Were Described?	If Validation Procedures (Cross-Validation, Independent Test Sets) Were Clearly Stated?
[19]	Low Risk	Low Risk	High Risk
[20]	Low Risk	Low Risk	Low Risk
[21]	Low Risk	Low Risk	High Risk
[22]	Low Risk	High Risk	Low Risk
[23]	Low Risk	Low Risk	High Risk
[24]	Low Risk	Low Risk	Low Risk
[25]	Low Risk	Low Risk	Low Risk
[26]	Low Risk	Low Risk	Low Risk
[27]	Low Risk	Low Risk	Low Risk
[28]	Low Risk	Low Risk	Low Risk
[29]	Low Risk	High Risk	High Risk
[30]	High Risk	Low Risk	Low Risk
[31]	Low Risk	Low Risk	Low Risk
[32]	Low Risk	Low Risk	Unknown Risk
[33]	Low Risk	Low Risk	Low Risk
[34]	Low Risk	Low Risk	Low Risk
[35]	Low Risk	High Risk	Low Risk
[36]	Low Risk	Low Risk	Low Risk
[37]	Low Risk	Low Risk	Unknown Risk
[38]	Low Risk	Low Risk	Low Risk
[39]	Low Risk	Low Risk	Low Risk
[40]	Low Risk	Low Risk	Low Risk
[41]	Low Risk	Low Risk	Unknown Risk
[42]	Low Risk	Low Risk	Low Risk
[43]	Low Risk	Low Risk	Low Risk
[44]	Low Risk	Low Risk	Low Risk
[45]	Low Risk	Low Risk	Low Risk

.

For each domain, the bias is categorized into the following:

• Low risk: Clearly described and well-defined methods, consistent procedures.
• High risk: Methods unclear, inconsistent, or potentially biased
• Unclear risk: Insufficient details to assess risk confidently.

4.1. Chlorophyll Content Estimation

Chlorophyll is a critical compound in leafy greens, playing an essential role in the process of photosynthesis by absorbing light energy and converting it into chemical energy for the plant [46]. It directly influences plant health, development, and vigor while also serving as a key indicator of nutritional quality [47]. Moreover, chlorophyll content correlates with the levels of other important compounds such as nitrogen, making it a valuable metric for assessing plant health and productivity [47].

Both hyperspectral imaging systems and spectrometers operate between 305 nm and 2500 nm, with the most common range used in all studies being 400 nm to 1000 nm. Those setups enabled the researchers to capture the detailed spectral signatures associated with chlorophyll content. In spinach, a spectrometer demonstrated significant utility for capturing chlorophyll-related reflectance. Additionally, high-throughput hyperspectral analysis techniques for lettuce utilized advanced spectral configurations, offering comprehensive data for chlorophyll quantification and enabling precise characterization of chlorophyll variations. Table 3, Table 4, Table 5 and Table 6, contain comprehensive overview of the studies.

A variety of machine learning and statistical models were applied to process spectral data and predict chlorophyll content. Techniques included Partial Least Squares Regression (PLSR), Principal Component Analysis (PCA), Multiple Linear Regression (MLR), and advanced methods like Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs). For baby leaf lettuce (Table 4), models such as PLSR and PCR achieved high accuracy, with R-values of 0.9099 and 0.9094, respectively, and normalized RMSE values as low as 0.4147. CARS-PLSR was notably effective for spinach, achieving an R² of 0.89 and an RMSEP of 0.10. Similarly, automated machine learning techniques (e.g., XGBoost version not specified [30]) used with hyperspectral data demonstrated superior performance for lettuce, yielding an R² of 0.93.

4.2. Carotenoids Content Estimation

Carotenoids play a crucial role in lettuce and basil, which contributes to plant physiology but also human nutrition [48,49]. In plants, they are integral to photosynthesis, assisting in the process of light absorption and energy transfer while additionally protecting chlorophyll from oxidative damage caused by extensive light [49]. The protection of plants from stress conditions, such as high light intensity or environmental fluctuations, is provided by carotenoids such as lutein and beta carotene [49,50]. They are connected to the overall plant health and vigor, which makes them a reliable marker of stress and metabolic activities. Carotenoids, from a nutritional perspective, are valuable as precursors to vitamin A, and they possess antioxidant properties, which can contribute to human health [49]. Leafy greens such as lettuce, spinach, and basil are a significant dietary contributor to carotenoids consumption [51,52].

In the articles collected, there is a wide range of spectral technologies, including both hyperspectral imaging systems and spectrometers. Hyperspectral imaging systems cover ranges from 350 nm to 2500 nm. These systems provided detailed spectral signatures associated with carotenoid content. Spectrometers operated between 340 nm and 900 nm. The most common range used in all studies is 400 nm to 1000 nm. These setups enabled the researchers to capture the detailed spectral signatures associated with carotenoid content. For instance, in baby leaf lettuce, a hyperspectral camera operating within 350 to 2500 nm (Table 7 and Table 8) demonstrated significant utility for capturing carotenoids related reflectance. Additionally, high-throughput hyperspectral analysis techniques for lettuce utilized advanced spectral configurations, offering comprehensive data for carotenoids quantification and enabling precise characterization of carotenoid variations.

Hyperspectral and spectroscopy combined with machine learning have proven to be a powerful tool for the detection of carotenoids. Across presented studies, accuracies were reported with values R² exceeding 0.85 and RMSE values of 0.03. Used methods include Partial Least Squares Regression (PLSR), Principal Component Regression (PCR), and Artificial Neural Networks (ANNs), with multitask learning models like MTCNN used as well. In some cases, advanced multitask learning models showed strong performance, particularly for spinach, achieving R² values above 0.8644 (Table 7) in VNIR and 0.8167 in NIR ranges.

Table 3. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce (baby leaf)	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Mulitvariate PLS R = 0.9099 NRMSE = 0.4147. Multivariate PCR R = 0.9094 NRMSE = 0.4159. Multivariate ANN with normalized full HSI data R = 0.8424 NRMSE = 0.3737.	300	[22]
Spinach	Dual-channel diode array spectrometer: 305 nm–1800 nm	Data dimensionality: Three methods: Competitive adaptive reweighed sampling (CARS) was used to eliminate the irrelevant variables of PLSR. VIs: PRI, PRI 500, PRI norm, RDVI, VOGREI3, NDVI, mrNDVI, TCARI, REIP, WI calculated from the spectrum Parameters based on inflection points (IPs) of the spectrum (first and second derivatives)	CARS-PLSR R2 = 0.89 RMSEP = 0.10 Best VI linear model: REIP R2 = 0.86 p < 0.001 Best IP linear model: IP1 R2 = 0.86 p < 0.001	1120	[27]
Lettuce	Fiberoptic spectrometer: 350 nm–2500 nm RGB Camera	Data dimensionality: Spectral vegetation indices (SVI) and color vegetation indices (CVI) were calculated from the hyperspectral data and used as inputs for the prediction models.	AutoML (XGBoost, All Indices) R2 0.93 RMSE N/A AutoML (GRVI) R2 0.91 RMSE 5.50 AutoML (CIgreen) R2 0.90 RMSE 6.15 AutoML (MCARI) R2 0.88 RMSE 6.43 AutoML (NDVI) R2 0.89 RMSE 6.20 Random Forest (All Indices) R2 0.91 RMSE N/A Random Forest (GRVI) R2 0.89 RMSE 5.81 Random Forest (NDVI) R2 0.78 RMSE 8.00 PLSR (GRVI) R2 0.82 RMSE 7.40 PLSR (NDVI) R2 0.79 RMSE 7.44 BPNN (CIgreen) R2 0.87 RMSE 6.23 BPNN (NDVI) R2 0.84 RMSE 6.83 SVM (GRVI) R2 0.81 RMSE 8.81 SVM (NDVI) R2 0.77 RMSE 7.94	3600 fiberoptic measurement 800 RGB	[30]
Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: - First derivative of the reflectance to identify band regions with the most differences - PLSR and PCA to identify the bands with more weight - VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.95 RMSE = 1.05 PLSR/PCA: Rp2 = 0.94 RMSE = 1.171 VIP-Score Index Rp2 = 0.99 RMSE = 0.03 However, there is high performance variability depending on the dataset	Not specified	[37]

Table 3. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce (baby leaf)	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Mulitvariate PLS R = 0.9099 NRMSE = 0.4147. Multivariate PCR R = 0.9094 NRMSE = 0.4159. Multivariate ANN with normalized full HSI data R = 0.8424 NRMSE = 0.3737.	300	[22]
Spinach	Dual-channel diode array spectrometer: 305 nm–1800 nm	Data dimensionality: Three methods: Competitive adaptive reweighed sampling (CARS) was used to eliminate the irrelevant variables of PLSR. VIs: PRI, PRI 500, PRI norm, RDVI, VOGREI3, NDVI, mrNDVI, TCARI, REIP, WI calculated from the spectrum Parameters based on inflection points (IPs) of the spectrum (first and second derivatives)	CARS-PLSR R2 = 0.89 RMSEP = 0.10 Best VI linear model: REIP R2 = 0.86 p < 0.001 Best IP linear model: IP1 R2 = 0.86 p < 0.001	1120	[27]
Lettuce	Fiberoptic spectrometer: 350 nm–2500 nm RGB Camera	Data dimensionality: Spectral vegetation indices (SVI) and color vegetation indices (CVI) were calculated from the hyperspectral data and used as inputs for the prediction models.	AutoML (XGBoost, All Indices) R2 0.93 RMSE N/A AutoML (GRVI) R2 0.91 RMSE 5.50 AutoML (CIgreen) R2 0.90 RMSE 6.15 AutoML (MCARI) R2 0.88 RMSE 6.43 AutoML (NDVI) R2 0.89 RMSE 6.20 Random Forest (All Indices) R2 0.91 RMSE N/A Random Forest (GRVI) R2 0.89 RMSE 5.81 Random Forest (NDVI) R2 0.78 RMSE 8.00 PLSR (GRVI) R2 0.82 RMSE 7.40 PLSR (NDVI) R2 0.79 RMSE 7.44 BPNN (CIgreen) R2 0.87 RMSE 6.23 BPNN (NDVI) R2 0.84 RMSE 6.83 SVM (GRVI) R2 0.81 RMSE 8.81 SVM (NDVI) R2 0.77 RMSE 7.94	3600 fiberoptic measurement 800 RGB	[30]
Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: - First derivative of the reflectance to identify band regions with the most differences - PLSR and PCA to identify the bands with more weight - VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.95 RMSE = 1.05 PLSR/PCA: Rp2 = 0.94 RMSE = 1.171 VIP-Score Index Rp2 = 0.99 RMSE = 0.03 However, there is high performance variability depending on the dataset	Not specified	[37]

Table 4. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral Camera: 400 nm–1000 nm	Data dimensionality: Spectral index (CI700) from the HRI	Does not provide specific accuracy metrics. It discusses the relationships between hyperspectral indices (CI700, PSRI) and physiological indicators of senescence	Not specified	[21]
Lettuce	Double-beam spectrophotometer: 340 nm–900 nm.	Not specified	Correlation between the measured reflectance and transmittance spectrum and the one estimated from PROSPECT-D model, which uses pigment quantities as input to see if a model inversion is possible. Only a graph is given for the result.	Not specified	[23]
Spinach	Hyperspectral camera: 400 nm–1000 nm 900 nm–1700 nm	Not specified	Unpackaged Spinach Leaves Best performing model Singletask regression model VNIR: BPNN Chla R2 0.8263 RMSE 0.2321 CNN Chlb R2 0.8527 RMSE 0.1296 CNN Chlt R2 0.8053 RMSE 0.1560 NIR: CNN Chla R2 0.7525 RMSE 0.3319 CNN Chlb R2 0.8514 RMSE 0.1994 PLSR Chlt R2 0.8050 RMSE 0.3484 Multitask regression model Best performing model: MTCNN VNIR: Chla R2 0.8683 RMSE 0.2034 Chlb R2 0.8416 RMSE 0.1306 Chlt R2 0.8681 RMSE 0.3064 Best model: MTPLSR NIR: Chla R2 0.8046 RMSE 0.2268 Chlb R2 0.8123 RMSE 0.1291 Chlt R2 0.8188 RMSE 0.3426 Packaged Spinach Leaves Best performing model Singletask regression model VNIR: BPNN Chla R2 0.9616 RMSE 0.1827 BPNN Chlb R2 0.9358 RMSE 0.3229 BPNN Chlt R2 0.9576 RMSE 0.4516 NIR: PLSR Chla R2 0.8870 RMSE 0.2930 PLSR Chlb R2 0.8514 RMSE 0.1994 PLSR Chlt R2 0.8797 RMSE 0.4797 Multitask regression model Best performing model: MTPLSR VNIR: Chla R2 0.9452 RMSE 0.2852 Chlb R2 0.9221 RMSE 0.2062 Chlt R2 0.9396 RMSE 0.4856 Best model: MTPLSR NIR: Chla R2 0.8874 RMSE 0.3932 Chlb R2 0.8460 RMSE 0.2563 Chlt R2 0.8763 RMSE 0.6421	Unpackaged spinach leaves 120 pictures. Packaged spinach leaves 150 pictures	[24]

Table 4. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral Camera: 400 nm–1000 nm	Data dimensionality: Spectral index (CI700) from the HRI	Does not provide specific accuracy metrics. It discusses the relationships between hyperspectral indices (CI700, PSRI) and physiological indicators of senescence	Not specified	[21]
Lettuce	Double-beam spectrophotometer: 340 nm–900 nm.	Not specified	Correlation between the measured reflectance and transmittance spectrum and the one estimated from PROSPECT-D model, which uses pigment quantities as input to see if a model inversion is possible. Only a graph is given for the result.	Not specified	[23]
Spinach	Hyperspectral camera: 400 nm–1000 nm 900 nm–1700 nm	Not specified	Unpackaged Spinach Leaves Best performing model Singletask regression model VNIR: BPNN Chla R2 0.8263 RMSE 0.2321 CNN Chlb R2 0.8527 RMSE 0.1296 CNN Chlt R2 0.8053 RMSE 0.1560 NIR: CNN Chla R2 0.7525 RMSE 0.3319 CNN Chlb R2 0.8514 RMSE 0.1994 PLSR Chlt R2 0.8050 RMSE 0.3484 Multitask regression model Best performing model: MTCNN VNIR: Chla R2 0.8683 RMSE 0.2034 Chlb R2 0.8416 RMSE 0.1306 Chlt R2 0.8681 RMSE 0.3064 Best model: MTPLSR NIR: Chla R2 0.8046 RMSE 0.2268 Chlb R2 0.8123 RMSE 0.1291 Chlt R2 0.8188 RMSE 0.3426 Packaged Spinach Leaves Best performing model Singletask regression model VNIR: BPNN Chla R2 0.9616 RMSE 0.1827 BPNN Chlb R2 0.9358 RMSE 0.3229 BPNN Chlt R2 0.9576 RMSE 0.4516 NIR: PLSR Chla R2 0.8870 RMSE 0.2930 PLSR Chlb R2 0.8514 RMSE 0.1994 PLSR Chlt R2 0.8797 RMSE 0.4797 Multitask regression model Best performing model: MTPLSR VNIR: Chla R2 0.9452 RMSE 0.2852 Chlb R2 0.9221 RMSE 0.2062 Chlt R2 0.9396 RMSE 0.4856 Best model: MTPLSR NIR: Chla R2 0.8874 RMSE 0.3932 Chlb R2 0.8460 RMSE 0.2563 Chlt R2 0.8763 RMSE 0.6421	Unpackaged spinach leaves 120 pictures. Packaged spinach leaves 150 pictures	[24]

Table 5. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Spinach	Hyperspectral camera: 470–900 nm	Data dimensionality: Attention embedded model	Spectral-feature model: PLS R2 0.8078 RMSE 3.0282 RPD 2.2903 SVM R2 0.8308 RMSE 2.8408 RPD 2.4414 1D ResNet R2 0.8457 RMSE 2.7128 RPD 2.5566 Spatial-feature model 2D ResNet R2 0.8495 RMSE 2.6789 RPD 2.5889 Spectral-spatial-feature 3D CNN models: C3D R2 0.8672 RMSE 2.5170 RPD 2.7554 3D ResNet R2 0.8786 RMSE 2.4204 RPD 2.8654 3D SqueezeNet R2 0.8740 RMSE 2.4519 RPD 2.8286 3D MobileNet R2 0.8729 RMSE 2.4621 RPD 2.8169 Attention embedded model 3D ResNet + channel + band R2 0.8998 RMSE 2.1865 RPD 3.1719	720	[25]
Lettuce	Fiber optic spectrometer: 200 nm–1100 nm (400 nm–980 nm considered)	Data Dimensionality: Three spectral indices were used (Pg, NDVI705, Pg/Pr)	Regression analysis Pg R2 = 0.677, F-value = 12.569 NDVI705 R2 = 0.789, F-value = 22.407 Pg/Pr R2 = 0.755, F-value = 40.009	45	[26]
Basil	Hyperspectral camera: 470 nm–900 nm	Not specified	Preprocessing (SG, SG + MSC, SG + SNV, SG + VSN) combined with machine learning methods (PLS, SVM, RF) Best Results: RF-SG-SNV R2 calibration: 0.905, RMSEC: 0.094 R2 prediction: 0.852, RMSEP: 0.120, RPD: 2.614 RF-SG-MSC R2 calibration: 0.878, RMSEC: 0.106 R2 prediction: 0.838, RMSEP: 0.126, RPD: 2.480 RF-SG-VSN R2 calibration: 0.857, RMSEC: 0.115 R2 prediction: 0.841, RMSEP: 0.124, RPD: 2.536	324	[28]
Basil	Hyperspectral camera: 470 nm–900 nm	Data dimensionality: The most important bands were selected by using an attention mechanism-based module in front of a 3D CNN (3D ResNet). For comparison, other band selection methods were used (SPA, GA, 2B-CNN).	The proposed model (attention band + 3D ResNet) was tested against other machine learning methods (SVM, 1D-CNN, 2D-CNN, 3D ResNet) using the reduced data (all method combinations). Best results obtained with the proposed model using the attention-based reduction: R2 = 0.912 RMSE = 2.046	540 hyperspectral images	[53]

Table 5. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Spinach	Hyperspectral camera: 470–900 nm	Data dimensionality: Attention embedded model	Spectral-feature model: PLS R2 0.8078 RMSE 3.0282 RPD 2.2903 SVM R2 0.8308 RMSE 2.8408 RPD 2.4414 1D ResNet R2 0.8457 RMSE 2.7128 RPD 2.5566 Spatial-feature model 2D ResNet R2 0.8495 RMSE 2.6789 RPD 2.5889 Spectral-spatial-feature 3D CNN models: C3D R2 0.8672 RMSE 2.5170 RPD 2.7554 3D ResNet R2 0.8786 RMSE 2.4204 RPD 2.8654 3D SqueezeNet R2 0.8740 RMSE 2.4519 RPD 2.8286 3D MobileNet R2 0.8729 RMSE 2.4621 RPD 2.8169 Attention embedded model 3D ResNet + channel + band R2 0.8998 RMSE 2.1865 RPD 3.1719	720	[25]
Lettuce	Fiber optic spectrometer: 200 nm–1100 nm (400 nm–980 nm considered)	Data Dimensionality: Three spectral indices were used (Pg, NDVI705, Pg/Pr)	Regression analysis Pg R2 = 0.677, F-value = 12.569 NDVI705 R2 = 0.789, F-value = 22.407 Pg/Pr R2 = 0.755, F-value = 40.009	45	[26]
Basil	Hyperspectral camera: 470 nm–900 nm	Not specified	Preprocessing (SG, SG + MSC, SG + SNV, SG + VSN) combined with machine learning methods (PLS, SVM, RF) Best Results: RF-SG-SNV R2 calibration: 0.905, RMSEC: 0.094 R2 prediction: 0.852, RMSEP: 0.120, RPD: 2.614 RF-SG-MSC R2 calibration: 0.878, RMSEC: 0.106 R2 prediction: 0.838, RMSEP: 0.126, RPD: 2.480 RF-SG-VSN R2 calibration: 0.857, RMSEC: 0.115 R2 prediction: 0.841, RMSEP: 0.124, RPD: 2.536	324	[28]
Basil	Hyperspectral camera: 470 nm–900 nm	Data dimensionality: The most important bands were selected by using an attention mechanism-based module in front of a 3D CNN (3D ResNet). For comparison, other band selection methods were used (SPA, GA, 2B-CNN).	The proposed model (attention band + 3D ResNet) was tested against other machine learning methods (SVM, 1D-CNN, 2D-CNN, 3D ResNet) using the reduced data (all method combinations). Best results obtained with the proposed model using the attention-based reduction: R2 = 0.912 RMSE = 2.046	540 hyperspectral images	[53]

Table 6. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral camera: 387 nm–1003 nm (437 nm–919 nm considered)	Data dimensionality: An attention module was put in front of a 1D-CNN to identify the most important spectral bands SPA algorithm was used as comparison band selection method.	One-dimensional-CNN Attention Module R2: 0.762 RMSE: 1.868 PLSR (Full spectrum) R2: 0.771 RMSE: 2.107 RF (Full spectrum) R2: 0.781 RMSE: 2.19 SPA + FDR + PLSR: R2: 0.742 RMSE: 2.440 SPA + FDR + RF: R2: 0.714 RMSE: 2.567	478	[29]
Lettuce	Hand-held spectrometer: 400 nm–1000 nm	Data dimensionality: The most important spectral bands were selected using uninformative variable elimination (UVE) combined with PLS. Results were compared with 30 SVI (MDATT especially)	UVE-PLS: R2 0.834 RMSE 38.58 mg m−2 MDATT: R2 0.736 RMSE 48.54 mg m−2	90	[39]
Lettuce	Multispectral camera: 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR).	Usage of VIs	Correlation between the most significant VIs and chlorophyll content. Chlorophyll a BNDVI R 0.89 GNDVI R 0.90 NDVI R 0.90 Chlorophyll b BNDVI R 0.89 GNDVI R 0.90 NDVI R 0.90	Not specified	[41]
Basil	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Vegetative Indices (MND705, MND750/700) were calculated from the spectra. ANOVA analysis was used to see the significance of these parameters.	Weak correlation between the VIs (MND705, MND750/700) and total chlorophyll due to the uneven spectra of the purple basil. Results: MND705 R2 = 0.1822 MND750/700 R2 = 0.1539	Not specified	[19]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	Linear regression was performed for the significant RI Chl (a + b): ChlRI R2 = 0.910 p = 0.0002 Car/Chl SIPI R2 = 0.72 p = 0.0077	60	[31]
Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.97 RMSE = 2.34 PCA R2 = 0.96 RMSE = 2.53	28	[32]

Table 6. Chlorophyll.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral camera: 387 nm–1003 nm (437 nm–919 nm considered)	Data dimensionality: An attention module was put in front of a 1D-CNN to identify the most important spectral bands SPA algorithm was used as comparison band selection method.	One-dimensional-CNN Attention Module R2: 0.762 RMSE: 1.868 PLSR (Full spectrum) R2: 0.771 RMSE: 2.107 RF (Full spectrum) R2: 0.781 RMSE: 2.19 SPA + FDR + PLSR: R2: 0.742 RMSE: 2.440 SPA + FDR + RF: R2: 0.714 RMSE: 2.567	478	[29]
Lettuce	Hand-held spectrometer: 400 nm–1000 nm	Data dimensionality: The most important spectral bands were selected using uninformative variable elimination (UVE) combined with PLS. Results were compared with 30 SVI (MDATT especially)	UVE-PLS: R2 0.834 RMSE 38.58 mg m−2 MDATT: R2 0.736 RMSE 48.54 mg m−2	90	[39]
Lettuce	Multispectral camera: 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR).	Usage of VIs	Correlation between the most significant VIs and chlorophyll content. Chlorophyll a BNDVI R 0.89 GNDVI R 0.90 NDVI R 0.90 Chlorophyll b BNDVI R 0.89 GNDVI R 0.90 NDVI R 0.90	Not specified	[41]
Basil	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Vegetative Indices (MND705, MND750/700) were calculated from the spectra. ANOVA analysis was used to see the significance of these parameters.	Weak correlation between the VIs (MND705, MND750/700) and total chlorophyll due to the uneven spectra of the purple basil. Results: MND705 R2 = 0.1822 MND750/700 R2 = 0.1539	Not specified	[19]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	Linear regression was performed for the significant RI Chl (a + b): ChlRI R2 = 0.910 p = 0.0002 Car/Chl SIPI R2 = 0.72 p = 0.0077	60	[31]
Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.97 RMSE = 2.34 PCA R2 = 0.96 RMSE = 2.53	28	[32]

Table 7. Carotenoids.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.655 NRMSE = 0.4254 Multivariate PCR R = 0.6229 NRMSE = 0.4301 Multivariate ANN with normalized full HSI data R = 0.7857 NRMSE = 0.611	300	[22]
Basil	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Vegetative Index (CRI700) was calculated from the spectra. ANOVA analysis was used to see the significance of this parameter.	Weak correlation for CRI700 due to the shielding effect of anthocyanins. Results: CRI700 R2 = 0.0479	Not specified	[19]
Lettuce	Hyperspectral Camera: 400 nm–1000 nm	Data dimensionality: Spectral index (PSRI) from the HRI	The document does not provide specific accuracy metrics. It discusses the relationships between hyperspectral indices (CI700, PSRI) and physiological indicators of senescence	Not specified	[21]
Lettuce	Double-beam spectrophotometer: 340 nm–900 nm.	Not specified	Correlation between the measured reflectance and transmittance spectrum and the one estimated from PROSPECT-D model, which uses pigment quantities as input to see if a model inversion is possible. Only a graph is given for the result.	Not specified	[23]
Spinach	Hyperspectral camera: 400 nm–1000 nm 900 nm–1700 nm	Not specified	Unpackaged Spinach Leaves Best performing model Single-task regression model VNIR: BPNN Car R2 0.7110 RMSE 0.2482 NIR: CNN Car R2 0.7523 RMSE 0.0348 Multitask regression model MTCNN VNIR: Car R2 0.8133 RMSE 0.0397 MTPLSR NIR: Car R2 0.7375 RMSE 0.0360 Packaged Spinach Leaves Best performing model Single-task regression model VNIR: BPNN Car R2 0.8436 RMSE 0.0802 NIR: PLSR Car R2 0.7619 RMSE 0.0565 Multitask regression model Best performing model: MTPLSR VNIR: Car R2 0.8644 RMSE 0.0426 MTPLSR NIR: Car R2 0.8167 RMSE 0.0552	Unpackaged spinach leaves 120 pictures per camera Packaged spinach leaves 150 pictures per camera	[24]

Table 7. Carotenoids.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.655 NRMSE = 0.4254 Multivariate PCR R = 0.6229 NRMSE = 0.4301 Multivariate ANN with normalized full HSI data R = 0.7857 NRMSE = 0.611	300	[22]
Basil	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Vegetative Index (CRI700) was calculated from the spectra. ANOVA analysis was used to see the significance of this parameter.	Weak correlation for CRI700 due to the shielding effect of anthocyanins. Results: CRI700 R2 = 0.0479	Not specified	[19]
Lettuce	Hyperspectral Camera: 400 nm–1000 nm	Data dimensionality: Spectral index (PSRI) from the HRI	The document does not provide specific accuracy metrics. It discusses the relationships between hyperspectral indices (CI700, PSRI) and physiological indicators of senescence	Not specified	[21]
Lettuce	Double-beam spectrophotometer: 340 nm–900 nm.	Not specified	Correlation between the measured reflectance and transmittance spectrum and the one estimated from PROSPECT-D model, which uses pigment quantities as input to see if a model inversion is possible. Only a graph is given for the result.	Not specified	[23]
Spinach	Hyperspectral camera: 400 nm–1000 nm 900 nm–1700 nm	Not specified	Unpackaged Spinach Leaves Best performing model Single-task regression model VNIR: BPNN Car R2 0.7110 RMSE 0.2482 NIR: CNN Car R2 0.7523 RMSE 0.0348 Multitask regression model MTCNN VNIR: Car R2 0.8133 RMSE 0.0397 MTPLSR NIR: Car R2 0.7375 RMSE 0.0360 Packaged Spinach Leaves Best performing model Single-task regression model VNIR: BPNN Car R2 0.8436 RMSE 0.0802 NIR: PLSR Car R2 0.7619 RMSE 0.0565 Multitask regression model Best performing model: MTPLSR VNIR: Car R2 0.8644 RMSE 0.0426 MTPLSR NIR: Car R2 0.8167 RMSE 0.0552	Unpackaged spinach leaves 120 pictures per camera Packaged spinach leaves 150 pictures per camera	[24]

Table 8. Carotenoids.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Multispectral camera: 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR).	Usage of VIs	Correlation between the most significant VIs and carotenoid content. BNDVI R 0.90 GNDVI R 0.89 NDVI R 0.90	Not specified	[41]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	Linear regression was performed for the significant RI Car/Chl SIPI R2 = 0.72 p = 0.0077	60	[31]

Table 8. Carotenoids.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Multispectral camera: 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR).	Usage of VIs	Correlation between the most significant VIs and carotenoid content. BNDVI R 0.90 GNDVI R 0.89 NDVI R 0.90	Not specified	[41]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	Linear regression was performed for the significant RI Car/Chl SIPI R2 = 0.72 p = 0.0077	60	[31]

4.3. Anthocyanin Content Estimation

Anthocyanins are a photochemical water water-soluble pigment from the flavonoid group. They are responsible for the fruit and vegetables color such as red, purple, and blue [54]. These pigments play an important role in plants, such as protection from ultraviolet radiation and oxidative stress, but also for attracting pollinators [55]. In vegetables such as lettuce, basil, or spinach, anthocyanins are responsible for color and stress tolerance [56]. When the outer leaves are under excessive light, these compounds act as a sunscreen by shielding the chloroplast [57]. This is clearly visible in the purple-leafed varieties of lettuce and basil. Anthocyanins are also connected to the plant’s response to the environment, such as light, temperature, and nutrients. In basil, it can be observed that anthocyanin levels can rise under strong UV exposure to enhance plant resistance [57]. In lettuce with red leaf varieties, the amount of anthocyanins is higher, which can correlate with higher antioxidant capacity [56]. In this study, various spectral imaging and spectrometry techniques were evaluated for their effectiveness in non-destructive anthocyanins detection in basil and lettuce, while no articles were found on non-destructive anthocyanins detection in spinach. These articles collected a wide range of spectral technologies, including both hyperspectral imaging systems, multispectral systems, and spectrometers without accuracy. Hyperspectral images were taken in 400 nm–1000 nm range, multispectral images were taken in 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR) wavelengths and spectrometer was working in 340 nm–1000 nm range. Those ranges enabled anthocyanins estimation with hyperspectral cameras and multispectral cameras. The best workflow in the article utilized UVE + CARS + SNV + DBO + ELM approach that resulted in R² = 0.8617, RMSE = 0.0095, RPD (Test Set): 2.7192 [20].

Table 9. Anthocyanins.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Basil	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Vegetative Indices (ARI, mARI) were calculated from the spectra. ANOVA analysis was used to see the significance of these parameters.	High correlation between VIs (ARI, mARI) and anthocyanins. Results: ARI R2 = 0.82 mARI R2 = 0.85	Not specified	[19]
Lettuce	Hyperspectral Camera: 400 nm–1000 nm	Data dimensionality: Spectral index (ARI) from the HRI	The document does not provide specific accuracy metrics. It discusses the relationships between hyperspectral indices (e.g., CI700, PSRI, and ARI) and physiological indicators of senescence	Not specified	[21]
Lettuce	Multispectral camera: 475 nm (blue), 560 nm (green), 668 nm (red), 717 nm (red edge), and 840 nm (NIR).	Usage of VIs	Correlation between the most significant VIs and anthocyanin content BNDVI R 0.83 GNDVI R 0.81 NDVI R 0.82	Not specified	[41]
Lettuce	Hyperspectral Camera: 400–1000 nm	Data Dimensionality Reduction: UVE and UVE + CARS were used to eliminate uninformative variables and focus on critical wavelengths.	Model 1 data on test UVE + CARS + SNV + DBO + ELM R2 = 0.8617, RMSE = 0.0095, RPD (Test Set): 2.7192 UVE + CARS + FD + SABO + ELM R2 0.8255, RMSE 0.010, RPD 2.4207 UVE + SNV + DBO + ELM R2 0.7986, RMSE 0.0123, RPD 2.2533 VI3 + WOA + ELM R2 0.812, RMSE 0.011, RPD 2.3323 UVE + Raw + DBO + ELM R2 0.7612, RMSE 0.0142, RPD 2.0693	135 hyperspectral picture	[20]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	Linear regression was performed for the significant RI Anthocyanins ARI R2 = 0.57 p = 0.029	60	[31]

, contain comprehensive overview of the studies about anthocyanin.

4.4. Nitrogen Nitrate and Nitrite Content Estimation

Nitrogen is one of the most important macro nutrients for plants; it directly influences growth, productivity, and development. Nitrogen plays an essential role in the formation of chlorophyll and the synthesis of amino acids, nucleic acids, and proteins in lettuce, spinach, and basil [58]. Nitrogen absorbance is usually in the form of nitrate ${\mathrm{NO}}_3^{-}$ and ammonium ${\mathrm{NH}}_4^{+}$. For most agricultural systems, nitrate is a nutrient and a signaling molecule that controls plant growth and metabolism. During the reduction in nitrate to ammonium, nitrite ${\mathrm{NO}}_2^{-}$ can form, and while the levels are usually low, this compound is important in nitrogen assimilation and metabolic pathways [58]. Once absorbed, nitrate is reduced to nitrite by the enzyme nitrate reductase. Nitrite is then reduced to ammonium by another enzyme called nitrite reductase. Ammonium helps with plant growth by being assimilated into amino acids and protein [58]. This part of the process needs light and temperature because it is energy dependent [58]. The challenge is balancing nitrogen, nitrate, and nitrite to optimize plant growth and health and minimize human consumption risks [59]. Nitrate serves as a storage of nitrogen and supports growth, but too much accumulation of it in leafy greens can be a health concern due to potential conversion into nitrite and nitrosamines in the human body, which have carcinogenic effects [59]. No articles were found for non-destructive nitrite detection in spinach, lettuce, or basil. Also, there are no articles for the estimation of nitrogen and nitrate in basil. As can be seen from

Table 10. Nitrogen.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.8216 NRMES = 0.5701 Multivariate PCR R = 0.8198 NRMSE = 0.5727 Multivariate ANN with normalized full HSI data R = 0.4343 NRMES = 0.5356	300	[22]
Spinach	Dual-channel diode array spectrometer: 305 nm–1800 nm	Data dimensionality: Three methods: Competitive adaptive reweighted sampling (CARS) was used to eliminate the irrelevant variables of PLSR. VIs (PRI, PRI 500, PRI norm, RDVI, VOGREI3, NDVI, mrNDVI, TCARI, REIP, WI) calculated form the spectrum Parameters based on inflection points (IPs) of the spectrum (first and second derivatives)	CARS-PLSR R2 = 0.58 RMSEP = 0.35 Best VI linear model: mrNDVI R2 = 0.47 p < 0.001 Best IP linear model IP1 R2 = 0.44 p < 0.001 CARS-PLSR No data Best VI linear model: mrNDVI R2 = 0.16 p < 0.001 Best IP linear model IP5 R2 = 0.17 p < 0.001	1120 Spectrometer measurements	[27]
Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	The variation in the RI caused by nitrogen deficiency was assessed during a 21-day cultivation period: ${\eta }^2$ is the effect size of a factor in percent; p is the level of significance of the factor effect. Lettuce 1 (Vitaminnyi) ChlRI ${\eta }^2$ 52.7 p < 0.0001 SIPI ${\eta }^2$ 40.2 p < 0.0001 R800 ${\eta }^2$ 10.1 p = 0.0056 PRI ${\eta }^2$ 18.8 p = 0.0001 ARI ${\eta }^2$ 57.2 p < 0.0001 Lettuce 2 (Kokarda) ChlRI ${\eta }^2$ 50.9 p < 0.001 SIPI ${\eta }^2$ 0.02 p = 0.893 R800 ${\eta }^2$ 3.5 p = 0.091 PRI ${\eta }^2$ 9.7 p = 0.005 ARI ${\eta }^2$ 0.3 p = 0.865	60	[31]

,

Table 11. Nitrogen.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Harvest 1 (four weeks) Prediction of nitrogen PLSR Models PLS1 (single nutrient output): Rp2 = 0.72 RMSE = 0.58 PLS2 (multiple nutrient output): Rp2 = 0.65 RMSE = 0.64 Classification of nitrogen deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Best Results: Threshold 1: PLS1 F1 = 0.75 Precision = 0.67 Recall = 0.86 PLSDA F1 = 0.75 Precision = 0.67 Recall = 0.86 MLP F1 = 0.75 Precision = 0.67 Recall = 0.86 Threshold 2: PLS1 F1 = 0.86 Precision = 1 Recall = 0.75 Harvest 2 (six weeks) Prediction of nitrogen PLSR Models PLS1 (single nutrient output): Rp2 = 0.88 RMSE = 0.37 PLS2 (multiple nutrient output): Rp2 = 0.88 RMSE = 0.37 Classification of nitrogen deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Best Results: Threshold 1: PLS1 F1 = 1 Precision = 1 Recall = 1 PLS2 F1 = 1 Precision = 1 Recall = 1 PLSDA F1 = 1 Precision = 1 Recall = 1 MLP F1 = 1 Precision = 1 Recall = 1 Threshold 2: PLS1 F1 = 1 Precision = 1 Recall = 1 PLS2 F1 = 1 Precision = 1 Recall = 1 PLSDA F1 = 1 Precision = 1 Recall = 1 MLP F1 = 1 Precision = 1 Recall = 1	288	[38]
Lettuce	Portable spectroradiometer: 350 nm–2500 nm	Data Dimensionality: Three methods to select the optimal wavelengths: - Principal Component Analysis (PCA) - Genetic Algorithm (GA) - Sequential Forward Selection (SFS).	Three regression methods (PLSR, BPNN, RF) were tested to estimate the nitrogen content: PLSR + PCA R2 0.58 RMSE 1.14 GA R2 0.95 RMSE 0.42 SFS R2 0.96 RMSE 0.30 BPNN + PCA R2 0.87 RMSE 0.6 GA R2 0.85 RMSE 3.25 SFS R2 0.97 RMSE 0.25 RF + PCA R2 0.90 RMSE 0.55 GA R2 0.95 RMSE 0.39 SFS R2 0.96 RMSE 0.36	2304 spectra (6 measurements per leaf, 3 leaves per plant, 8 plants per group, 4 groups, 4 growth stages)	[44]

,

Table 12. Nitrate.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.98 RMSE = 346 PCA R2 = 0.97 RMSE = 455	28	[32]
Spinach	Spectroscopy 834 nm–1475 nm 2403 nm–2502 nm (noise regions were removed).	Data Dimensionality: Principal Component Analysis (PCA) was applied to identify outliers and preprocess the data.	MPLS Model: R2 0.38–0.45 Standard Error of Prediction (SEP):  920 mg/kg. LOCAL Algorithm: R2 0.6 SEP: 758 mg/kg.	516 spinach plants	[43]
Spinach	Near-Infrared Spectroscopy (NIRS): 1600 nm–2400 nm.	Data Dimensionality: Principal Component Analysis (PCA) was used to decompose and compress the data matrix and detect outliers before calibration modeling.	MPLS Validation results: R2 0.41, RDP = 1.29 Partial Least Squares-Discriminant Analysis (PLS-DA) was used for classifying nitrate content with classification accuracies ranging from 73 percent to 85 percent for different classes.	128 spinach samples were analyzed	[42]
Lettuce	800 to 2500 nm Spectrometer	Data dimensionality: PCA was used to identify outliers among the spectra, and LDA was used to divide the spectra based on the lettuce variety. After the regression, PLS managed to identify nine components (spectral bands)	PLSR Model (only 191 samples by removing outliers) Square of determinant coefficient after cross-validation (Q2) = 0.90 RMSECV = 99 mg/kg	1330 spectra (266 lettuce heads, 5 spectra per plant)	[33]
Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: - First derivative of the reflectance to identify band regions with the most differences - PLSR and PCA to identify the bands with more weight - VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.93 RMSE = 437.19 PLSR/PCA: Rp2 = 0.98 RMSE = 185.63 VIP-Score: Rp2 = 0.99 RMSE = 111.51 However, there is high performance variability depending on the dataset	Not specified	[37]

,

Table 13. Nitrate.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Spinach	Vis-NIR spectrophotometer: 350 nm–2500 nm (400 nm–2250 nm considered)	Data dimensionality: the most significant bands were selected with PLS using the full spectrum as input. Then, the PLS model was tested again with only the selected wavelengths.	PLS model with full spectral range SG 125 R2 0.865 cross-validation RPD 1.459 cross-validation SG 127 R2 0.882 cross-validation RPD 1.559 cross-validation SNV R2 0.908 cross-validation RPD 1.768 cross-validation Baseline Correction R2 0.892 cross-validation RPD 1.629 cross-validation Detrending DT R2 0.897 cross-validation RPD 1.673 cross-validation MSC R2 0.908 cross-validation RPD 1.767 cross-validation Raw unprocessed R2 0.898 cross-validation RPD 1.679 cross-validation PLS model with selected wavelengths Raw data R2 0.869 cross-validation RPD 1.482 cross-validation	261 spectra (9 treatments, 29 leaves per treatment)	[34]
Spinach	NIR spectrophotometer: 908 nm–1676 nm	Data Dimensionality: Principal Component Analysis (PCA) was applied to study the population structure, and some bands were identified. However, the MLPS regression was performed using the full measured spectrum.	In the first part, cross-validation analysis was used to optimize the number of spectra to be taken at each production chain step: 1 spectrum for each plant in the first two stages and 2 spectra for each plant in the last. After that, an MLPS regression was used to estimate the nitrate content in each production stage: Field R2cv 0.59 RPDcv 1.55 Laboratory R2cv 0.52 RPDcv 1.45 After Washing R2cv 0.54 RPDcv 1.46	4000 spectra (77 plants; 5 spectra per plant during first two phases; 6 spectra per leaf; and between 4 and 10 spectra per plant during last phase)	[35]

and

Table 14. Nitrate.

Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Lettuce	Spectrograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Prediction based on hyperspectral data using three methods (CNN, PLSR, SVR): CNN Inception Module Stratified Sampling R2P = 0.9166 RMSE = 679.5709 Normal Sampling R2P = 0.9073 RMSE = 717.5676 Residual Module R2P = 0.9235 RMSE = 650.8244 Attention Module R2P = 0.9232 RMSE = 652.2112 Residual Module + Attention Module R2P = 0.9317 RMSE = 615.0037 PLSR R2P = 0.8782 RMSE = 821.2243 SVR R2P = 0.8942 RMSE = 801.7419 Prediction based on time series phenotypes combined with the best deep learning (Optimal-Spec) models based on hyperspectral data using four methods (TD, LSTM, GRU, BRNN): Optimal-Spec-TD R2P = 0.9435 RMSE = 559.2437 Optimal-Spec-LSTM R2P = 0.9354 RMSE = 598.1094 Optimal-Spec-GRU R2P = 0.9400 RMSE = 576.2199 Optimal-Spec-BRNN R2P = 0.9380 RMSE = 581.2459	433 spectra (560 plants: 160 Butter, 200 Head, 100 Leaf, and 100 Roman)	[36]
Spinach	MicroNIR™ 1700 spectral range 910–1676 nm and Matrix-F (FT-NIR-based) spectral range 834–2502 nm (after removing noisy regions)	Data dimensionality: MPLS, PCA, SNV, DT, Spectral preprocessing: derivative treatments 1, 5, 5, 1 and 2, 5, 5, 1	MicroNIR™ 1700 R2cv = 0.50 SECV = 633.73 mg/kg RPFcv = 1.41 Matrix-F R2cv = 0.44 SECV = 67,614 mg/kg RPFcv = 1.33	195 spinach plants. After removing outliers, calibration sets included 144–146 samples, and validation sets included 47 samples.	[45]

many methods in the topic of non-destructive nitrogen and nitrate detection reached strong correlations and low error margins, with some reaching R² = 0.98 RMSE = 346 [32]. This demonstrates the high level of reliability that is reached for the estimation of those compounds. Such performance suggests that hyperspectral imaging and spectroscopy are not only possible but effective tool for nitrogen and nitrate detection in lettuce and spinach. These techniques have the potential to substitute traditional chemical analysis, which is inherently destructive to the plant. Spectral measurements can offer fa aster and more efficient solution for nutrient monitoring, which can be performed over time on the same plant or leaf.

4.5. Additional Biochemical Traits Detected

While the primary focus of this systematic review was the detection of chlorophyll, carotenoids, nitrogen, nitrate, and anthocyanins, several of the articles have identified additional biochemical traits in lettuce, basil, and spinach. These traits can provide deeper insights into plant physiology, nutrient content, and safety, offering promising opportunities for further research. Below, in

Table 15. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Sugar Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Sucrose Multivariate PLS R = 0.7092 NRMSE = 0.8398 Multivariate PCR R = 0.5771 NRMSE = 0.9038 Multivariate ANN with normalized full HSI data R = −0.04568 NRMSE = 0.6889 Glucose Multivariate PLS R = 0.8402 NRMSE = 0.5422 Multivariate PCR R = 0.7967 NRMSE = 0.6017 Multivariate ANN with normalized full HSI data R = 0.02561 NRMSE = 0.6746 Fructose Multivariate PLS R = 0.8579 NRMSE = 0.538 Multivariate PCR R = 0.8155 NRMSE = 0.5788 Multivariate ANN with normalized full HSI data R = 0.3967 NRMSE = 0.5251	300	[22]
Sugar, Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: - First derivative of the reflectance to identify band regions with the most differences - PLSR and PCA to identify the bands with more weight - VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.98 RMSE = 0.36 PLSR/PCA: Rp2 = 0.99 RMSE = 0.21 VIP-Score: Rp2 = 0.99 RMSE = 0.03 However, there is high performance variability depending on the dataset	Not specified	[37]
Phosphorus, Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.9094 NRMSE = 0.7649 Multivariate PCR R = 0.8911 NRMSE = 0.4539 Multivariate ANN with normalized full HSI data R = 0.7155 NRMSE = 0.6155	300	[22]
Phosphorus, Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	The variation in the RI caused by phosphorus deficiency was assessed during a 21-day cultivation period: ${\eta }^2$ is the effect size of a factor in percent; p is the level of significance of the factor effect. Lettuce 1 (Vitaminnyi) ChlRI ${\eta }^2$ 37.7 p < 0.0001 SIPI ${\eta }^2$ 11.7 p = 0.0026 R800 ${\eta }^2$ 10.0 p = 0.0057 PRI ${\eta }^2$ 26.9 p < 0.0001 ARI ${\eta }^2$ 47.3 p < 0.0001 Lettuce 2 (Kokarda) ChlRI ${\eta }^2$ 32.1 p < 0.0001 SIPI ${\eta }^2$ 0.6 p = 0.473 R800 ${\eta }^2$ 12.8 p = 0.0012 PRI ${\eta }^2$ 12.0 p = 0.0018 ARI ${\eta }^2$ 2.1 p = 0.202	60	[31]

,

Table 16. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Phosphorus, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Only the best results are shown. Harvest 1 (four weeks) Prediction of phosphorus PLSR Models PLS1 (single nutrient output): Rp2 = 0.71 RMSE = 0.13 PLS2 (multiple nutrient output): Rp2 = 0.74 RMSE = 0.12 Classification of phosphorus deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Threshold 1: PLSDA F1 = 0.86 Precision = 0.86 Recall = 0.86 Threshold 2: PLS1 F1 = 0.86 Precision = 1 Recall = 0.75 PLS2 F1 = 0.86 Precision = 1 Recall = 0.75 Harvest 2 (six weeks) Prediction of phosphorus PLSR Models PLS1 (single nutrient output): Rp2 = 0.68 RMSE = 0.12 PLS2 (multiple nutrient output): Rp2 = 0.58 RMSE = 0.14 Classification of phosphorus deficiency (F1, Precision, Recall) Threshold 1: MLP F1 = 0.75 Precision = 0.75 Recall = 0.75 Threshold 2: MLP F1 = 0.86 Precision = 1 Recall = 0.75	288	[38]
Phosphorus, Lettuce	Portable spectroradiometer: 350 nm–2500 nm	Data Dimensionality: Three methods to select the optimal wavelengths: - Principal Component Analysis (PCA) - Genetic Algorithm (GA) - Sequential Forward Selection (SFS).	Three regression methods (PLSR, BPNN, RF) were tested to estimate the phosphorus content: PLSR + PCA R2 0. 52 RMSE 4.48 GA R2 0.88 RMSE 1.92 SFS R2 0.83 RMSE 4.4 BPNN + PCA R2 0.9 RMSE 0.26 GA R2 0.88 RMSE 3 SFS R2 0.57 RMSE 4.35 RF + PCA R2 0.94 RMSE 0.2 GA R2 0.89 RMSE 0.35 SFS R2 0.85 RMSE 0.54	2304 spectra (6 measurements per leaf; 3 leaves per plant; 8 plants per group; 4 groups; and 4 growth stages).	[44]
Potassium, Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.6442 NRMSE = 0.7649 Multivariate PCR R = 0.545 NRMSE = 0.8385 Multivariance ANN with normalized full HSI data R = 0.6155 NRMSE = 0.7667	300	[22]

,

Table 17. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Potassium Lettuce	Fiberoptic spectroradiometer: 360 nm–1000 nm	Data dimensionality: Reflectance Indices (ChlRI, SIPI, R800, PRI, ARI) were calculated	The variation in the RI caused by phosphorus deficiency was assessed during a 21-day cultivation period: ${\eta }^2$ is the effect size of a factor in percent, p is the level of significance of the factor effect. Lettuce 1 (Vitaminnyi) ChlRI ${\eta }^2$ 20.4 p < 0.0001 SIPI ${\eta }^2$ 3.6 p = 0.101 R800 ${\eta }^2$ 0.01 p = 0.93 PRI ${\eta }^2$ 12.0 p < 0.0023 ARI ${\eta }^2$ 17.9 p < 0.0002 Lettuce 2 (Kokarda) ChlRI ${\eta }^2$ 30.7 p < 0.0001 SIPI ${\eta }^2$ 0.6 p = 0.829 R800 ${\eta }^2$ 1.1 p = 0.362 PRI ${\eta }^2$ 5.2 p = 0.042 ARI ${\eta }^2$ 2.6 p = 0.157	60	[31]
Potassium, Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.97 RMSE = 131 PCA R2 = 0.94 RMSE = 191	28	[32]
Potassium, Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: - First derivative of the reflectance to identify band regions with the most differences - PLSR and PCA to identify the bands with more weight - VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.97 RMSE = 83.69 PLSR/PCA: Rp2 = 0.97 RMSE = 80.84 VIP-Score: Rp2 = 0.99 RMSE = 30.02 However, there is high performance variability depending on the dataset	Not specified	[37]

,

Table 18. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Potassium, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Harvest 1 (four weeks) Prediction of Potassium PLSR Models PLS1 (single nutrient output): Rp2 = 0.69 RMSE = 1.5 PLS2 (multiple nutrient output): Rp2 = 0.67 RMSE = 1.54 Classification of potassium deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Best results: Threshold 1: PLS1 F1 = 0.73 Precision = 0.67 Recall = 0.80 Threshold 2: PLS1 F1 = 0.67 Precision = 1 Recall = 0.50 PLS2 F1 = 0.67 Precision = 1 Recall = 0.50 PLSDA F1 = 0.67 Precision = 1 Recall = 0.50 MLP F1 = 0.67 Precision = 1 Recall = 0.50 Harvest 2 (six weeks) Prediction of Potassium PLSR Models PLS1 (single nutrient output): Rp2 = 0.38 RMSE = 1.95 PLS2 (multiple nutrient output): Rp2 = 0.42 RMSE = 1.87 Classification of potassium deficiency (F1, Precision, Recall) Best results: Threshold 1: PLS2	288	[38]
Potassium, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	F1 = 0.73 Precision = 0.8 Recall = 0.67 PLSDA F1 = 0.73 Precision = 0.8 Recall = 0.67 MLP F1 = 0.73 Precision = 0.8 Recall = 0.67 Threshold 2: PLS1 F1 = 0.4 Precision = 1 Recall = 0.25 PLSDA F1 = 0.4 Precision = 1 Recall = 0.25	288	[38]
Potassium, Lettuce	Portable spectroradiometer: 350 nm–2500 nm	Data Dimensionality: Three methods to select the optimal wavelengths: - Principal Component Analysis (PCA) - Genetic Algorithm (GA) - Sequential Forward Selection (SFS).	Three regression methods (PLSR, BPNN, RF) were tested to estimate the potassium content: PLSR + PCA R2 0. 93 RMSE 0.37 GA R2 0.95 RMSE 0.34 SFS R2 0.92 RMSE 0.38 BPNN + PCA R2 0.9 RMSE 0.5 GA R2 0.9 RMSE 0.53 SFS R2 0.94 RMSE 0.36 RF + PCA R2 0.96 RMSE 0.32 GA R2 0.88 RMSE 2.83 SFS R2 0.96 RMSE 0.35	2304 spectra (6 measurements per leaf; 3 leaves per plant; 8 plants per group; 4 groups; and 4 growth stages).	[44]

,

Table 19. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Vitamin C, Lettuce	Hyperspectral Camera: 350 nm–2500 nm	Data dimensionality: First-order derivative (FDR) of reflectance was used to find the 13 most important principal components.	Multivariate PLS R = 0.655 NRMSE = 0.5701 Multivariate PCR R = 0.6229 NRMSE = 0.4301 Multivariance ANN with normalized full HSI data R = 0.3259 NRMSE = 0.7478	300	[22]
Vitamin C, Spinach	Near-Infrared Spectroscopy (NIRS): 1600 nm–2400 nm.	Data Dimensionality: Principal Component Analysis (PCA) was used to decompose and compress the data matrix and detect outliers before calibration modeling.	MPLS Validation results: R2 0.33, RDP = 1.21	128 spinach samples were analyzed	[42]
Water (Relative Water Content, RWC), Spinach	Dual-channel diode array spectrometer: 305 nm–1800 nm	Data dimensionality: Three methods: - Competitive adaptive reweighted sampling (CARS) was used to eliminate the irrelevant variables of PLSR. - VIs (PRI, PRI 500, PRI norm, RDVI, VOGREI3, NDVI, mrNDVI, TCARI, REIP, WI) calculated from the spectrum - Parameters based on inflection points (IPs) of the spectrum	CARS-PLSR R2 = 0.30 RMSEP = 1.43 Best VI linear model: WI R2 = 0.16 p = 0.019 Best IP linear model: IP1 R2 = 0.05 p < 0.001	1120	[27]
Water Stress, Lettuce	Spectrograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Prediction of water stress based on time series phenotypes combined with the best deep learning (Optimal-Spec) models based on hyperspectral data Optimal-Spec Accuracy = 96.59 percent Optimal-Spec-TD Accuracy = 98.86 percent	433 spectra (560 plants: 160 butter; 200 head; 100 leaf; and 100 Roman)	[36]

,

Table 20. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Calcium, Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.91 RMSE = 43.3 PCA R2 = 0.86 RMSE = 54.5	28	[32]
Calcium, Lettuce	Specrograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Prediction based on hyperspectral data using three methods (CNN, PLSR, SVR): CNN Inception Module Stratified Sampling R2P = 0.8136 RMSE = 38.9377 Normal Sampling R2P = 0.8027 RMSE = 34.0846 Residual Module R2P = 0.8325 RMSE = 36.9115 Attention Module R2P = 0.8460 RMSE = 35.3928 Residual Module + Attention Module R2P = 0.8390 RMSE = 36.1864 PLSR R2P = 0.7608 RMSE = 44.1029 SVR R2P = 0.7215 RMSE = 49.9887 Prediction based on time series phenotypes combined with the best deep learning (Optimal-Spec) models based on hyperspectral data using four methods (TD, LSTM, GRU, BRNN):	433 spectra (560 plants: 160 Butter, 200 Head, 100 Leaf, and 100 Roman)	[36]
Calcium, Lettuce	Spectograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Optimal-Spec-TD R2P = 0.8675 RMSE = 32.8215 Optimal-Spec-LSTM R2P = 0.8716 RMSE = 32.3151 Optimal-Spec-GRU R2P = 0.8670 RMSE = 32.8901 Optimal-Spec-BRNN R2P = 0.8593 RMSE = 33.8303	433 spectra (560 plants: 160 Butter, 200 Head, 100 Leaf, and 100 Roman)	[36]

,

Table 21. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Calcium, Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: First derivative of the reflectance to identify band regions with the most differences PLSR and PCA to identify the bands with more weight VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.75 RMSE = 23.76 PLSR/PCA: Rp2 = 0.65 RMSE = 27.87 VIP-Score: Rp2 = 0.65 RMSE = 25.42	Not specified	[37]
Calcium, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Only the best results are shown. Harvest 1 (four weeks) Prediction of calcium PLSR Models PLS1 (single nutrient output): Rp2 = 0.12 RMSE = 0.4 PLS2 (multiple nutrient output): Rp2 = 0.24 RMSE = 0.37 Classification of calcium deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Threshold 1: PLS1 F1 = 0.73 Precision = 0.67 Recall = 0.80 Threshold 2: MLP F1 = 0.5 Precision = 0.5 Recall = 0.5 Harvest 2 (six weeks) Prediction of calcium PLSR Models PLS1 (single nutrient output): Rp2 = 0.07 RMSE = 0.52 PLS2 (multiple nutrient output): Rp2 = 0.11 RMSE = 0.52 Classification of calcium deficiency (F1, Precision, Recall) Threshold 1: None F1 = 0 Precision = 0 Recall = 0 Threshold 2: MLP F1 = 0.69 Precision = 0.69 Recall = 0.69	288	[38]
Soluble Solid Content, Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.95 RMSE = 1.46 PCA R2 = 0.88 RMSE = 2.25	28	[32]
Soluble Solid Content, Spinach	Near-Infrared Spectroscopy (NIRS): 1600 nm–2400 nm.	Data Dimensionality Reduction: Principal Component Analysis (PCA) was used to decompose and compress the data matrix and detect outliers before calibration modeling.	MPLS Validation results: R2 0.85, RDP = 2.54	128	[42]

,

Table 22. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Soluble Solid Content, Spinach	NIR spectrophotometer: 908 nm–1676 nm	Data Dimensionality: Principal Component Analysis (PCA) was applied to study the population structure, and some bands were identified. However, the MLPS regression was performed using the full measured spectrum.	In the first part, cross-validation analysis was used to optimize the number of spectra to be taken at each production chain step: 1 spectrum for each plant in the first two stages and 2 spectra for each plant in the last. After that, an MLPS regression was used to estimate the SSC content in each production stage: Field R2cv 0.55 RPDcv 1.55 Laboratory R2cv 0.60 RPDcv 1.66 After Washing R2cv 0.62 RPDcv 1.76	4000 spectra (77 plants; 5 spectra per plant during first two phases; 6 spectra per leaf; and between 4 and 10 spectra per plant during last phase)	[35]
Soluble Solid Content, Lettuce	Spectrograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Prediction based on hyperspectral data using three methods (CNN, PLSR, SVR): CNN Inception Module Stratified Sampling R2P = 0.8307 RMSE = 0.4629 Normal Sampling R2P = 0.8279 RMSE = 0.4030 Residual Module R2P = 0.8413 RMSE = 0.4481 Attention Module R2P = 0.8694 RMSE = 0.4065 Residual Module + Attention Module R2P = 0.8743 RMSE = 0.3989 PLSR R2P = 0.7913 RMSE = 0.5139 SVR R2P = 0.8023 RMSE = 0.5599 Prediction based on time series phenotypes combined with the best deep learning (Optimal-Spec) models based on hyperspectral data using four methods (TD, LSTM, GRU, BRNN): Optimal-Spec-TD R2P = 0.8743 RMSE = 0.3989 Optimal-Spec-LSTM R2P = 0.8764 RMSE = 0.3955 Optimal-Spec-GRU R2P = 0.8772 RMSE = 0.3941 Optimal-Spec-BRNN R2P = 0.8828 RMSE = 0.3851	433 spectra (560 plants: 160 butter, 200 head, 100 leaf, and 100 Roman)	[36]
Cadmium, Lettuce	Fiber optic spectrometer: 200 nm–1100 nm (400 nm–980 nm considered)	Data Dimensionality: Four spectral indices were used	Regression analysis SDr/SDy R2 = 0.872, F-value = 24.959 SDb/SDy R2 = 0.781, F-value = 13.041 (SDr − SDy)/(SDr + SDy) R2 = 0.792, F-value = 13.996 (SDb − SDy)/(SDb + SDy) R2 = 0.65, F-value = 6.8	45	[26]

,

Table 23. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
pH, Lettuce	Hyperspectral camera: 400 nm–1000 nm	Data dimensionality: Four linear regression fit models were used to identify the 24 most significant wave bands (out of 448). The inputs of these models were reflectance, the first derivative of the reflectance, and two spectral indices (RSI, NDSI). After that, the bands were reduced to 22 due to the very small improvement of the regression models using 24 bands.	PLSR R2 = 0.97 RMSE = 0.03 PCA R2 = 0.81 RMSE = 0.09	28	[32]
pH, Lettuce	Spectrograph + CCD camera: 400 nm–1000 nm	Data dimensionality: to reduce the computational load of the CNN, an attention mechanism was employed to focus on significant features of the spectra	Prediction based on hyperspectral data using three methods (CNN, PLSR, SVR): CNN Inception Module Stratified Sampling R2P = 0.9116 RMSE = 0.1944 Normal Sampling R2P = 0.9079 RMSE = 0.2072 Residual Module R2P = 0.9472 RMSE = 0.1502 Attention Module R2P = 0.9399 RMSE = 0.1603 Residual Module + Attention Module R2P = 0.9240 RMSE = 0.1804 PLSR R2P = 0.9240 RMSE = 0.1804 SVR R2P = 0.8753 RMSE = 0.2374 Prediction based on time series phenotypes combined with the best deep learning (Optimal-Spec) models based on hyperspectral data using four methods (TD, LSTM, GRU, BRNN): Optimal-Spec-TD R2P = 0.9583 RMSE = 0.1336 Optimal-Spec-LSTM R2P = 0.9420 RMSE = 0.1575 Optimal-Spec-GRU R2P = 0.9404 RMSE = 0.1597 Optimal-Spec-BRNN R2P = 0.9464 RMSE = 0.1514	433 spectra (560 plants: 160 Butter, 200 Head, 100 Leaf, and 100 Roman)	[36]

,

Table 24. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
pH, Lettuce	Hyperspectral camera: 390.57 nm–1008.6 nm	Data dimensionality: Three Methods: First derivative of the reflectance to identify band regions with the most differences PLSR and PCA to identify the bands with more weight VIP score of a predictor variable over “1”	Best results of each method over four lettuce species cultivated with two different systems each (8 datasets) Results for hydroponically grown black-seeded Simpson lettuce: FDR: Rp2 = 0.95 RMSE = 0.03 PLSR/PCA: Rp2 = 0.98 RMSE = 0.02 VIP-Score: Rp2 = 0.98 RMSE = 0.016 However, there is high performance variability depending on the dataset	Not specified	[37]
Magnesium, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Harvest 1 (four weeks) Prediction of magnesium PLSR Models PLS1 (single nutrient output): Rp2 = 0.34 RMSE = 0.34 PLS2 (multiple nutrient output): Rp2 = 0.28 RMSE = 0.36 Classification of magnesium deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Threshold 1: PLSDA F1 = 0.57 Precision = 1 Recall = 0.40 Threshold 2: PLS1 F1 = 0.74 Precision = 0.78 Recall = 0.7 Harvest 2 (six weeks) Prediction of magnesium PLSR Models PLS1 (single nutrient output): Rp2 = 0.07 RMSE = 0.37 PLS2 (multiple nutrient output): Rp2 = 0.09 RMSE = 0.36 Classification of magnesium deficiency (F1, Precision, Recall) Best results: Threshold 1: None F1 = 0 Precision = 0 Recall = 0 Threshold 2: PLSDA F1 = 0.62 Precision = 0.57 Recall = 0.67	288	[38]

and

Table 25. Additional Biochemical Traits.

Biochemical Traits and Plant	Spectral Range and Instrument	Dimensionality Reduction Techniques	Accuracy and Technique	Number of Samples	Ref.
Sulfur, Lettuce	Hyperspectral camera: 390 nm–1000 nm	Data dimensionality: Segmentation of pixels based on two VIs (NDVI, GI), PLSR used to identify the predictor variables	Harvest 1 (four weeks) Prediction of Sulfur PLSR Models PLS1 (single nutrient output): Rp2 = 0.6 RMSE = 0.04 PLS2 (multiple nutrient output): Rp2 = 0.64 RMSE = 0.04 Classification of sulfur deficiency (F1, Precision, Recall) (PLS1, PLS2, PLSDA, MLP) Best results: Threshold 1: MLP F1 = 0.84 Precision = 0.72 Recall = 1 Threshold 2: PLS1 F1 = 0.71 Precision = 0.83 Recall = 0.63 PLSDA F1 = 0.71 Precision = 0.83 Recall = 0.63 Harvest 2 (six weeks) Prediction of Sulfur PLSR Models PLS1 (single nutrient output): Rp2 = 0.27 RMSE = 0.04 PLS2 (multiple nutrient output): Rp2 = 0.33 RMSE = 0.04 Classification of magnesium deficiency (F1, Precision, Recall) Best results: Threshold 1: MLP F1 = 0.84 Precision = 0.73 Recall = 1 Threshold 2: MLP F1 = 0.75 Precision = 0.75 Recall = 0.75	288	[38]

, findings are summarized.

These biochemical traits hold an important role in plant physiology, quality, and nutritional value, and their detection with spectral measurements can help with optimizing plant growth, quality, and providing insight into plant health.

• Glucose and sugar are energy storage and indicators of plant photosynthetic efficiency and are responsible for the sweetness and flavor profile of leafy greens [60]. Spectral measurements allow non-destructive quantification [37], which could improve understanding of metabolic processes and optimize quality traits in lettuce, basil, and spinach.
• Phosphorus is an important micronutrient needed by plants in large quantities. It helps with energy transfer through ATP (adenosine triphosphate) and with the formation of nucleic acid and cell membranes. It is essential in various tasks, such as plant growth, root development, and photosynthesis [61]. Adequate phosphorus levels are needed for maintaining metabolic balance and optimizing yield [61]. A non-destructive monitoring can lead to more precise management of nutrient inputs, which would lead to more efficient application of fertilizers. By doing so, costs and environmental impacts can be reduced.
• Potassium is a micronutrient that plays a role in regulating water balance in plants, activates enzymes, and has a significant role in photosynthesis [61]. It plays a crucial role in stomatal regulation, which helps plants efficiently control water loss and gas exchange; these processes are essential for maintaining plant hydration and nutrient transport. In leafy greens such as basil, lettuce, and spinach, potassium supports plant vigor, improves resistance to stress [61].
• Vitamin C (ascorbic acid) is a powerful antioxidant in plants; it is responsible for protecting them against oxidative stress that can be caused by excessive lighting, drought, and pollution [62]. By aiding in detoxification of reactive oxygen and maintaining cellular health, vitamin C is crucial for the plant defense mechanism [62,63]. In leafy greens such as lettuce, spinach, and basil, ascorbic acid contributes to the overall well-being of the plant and enhances the nutritional quality [62]. For plant development, it helps with enzymatic processes and growth regulations [62].
• Water Content is essential to determine plant hydration and freshness. Water supports photosynthesis and transports nutrients throughout the plant [64]. Insufficient amounts of water can lead to wilting, weak plant health, and reduced growth [65]. Excessive water can affect root functions and make the plant more susceptible to diseases [66].
• Calcium is a plant nutrient required in cell wall structure and membranes, which gives tissue integrity and strength [67]. Adequate calcium levels prevent physiological disorders and improve the structural quality of leaves, ensuring durability and freshness [61]. Calcium also contributes to the general resilience of plant tissues, helping them withstand mechanical stress [61].
• Soluble solid content (SSC) refers to the concentration of dissolved solids in plant tissues, primarily sugars, organic acids, and other soluble compounds
• Magnesium is a key component of chlorophyll and a cofactor for many enzymatic reactions [61].
• Sulfur is pivotal for protein synthesis, enzyme activity, and chlorophyll production [61].
• Cadmium and Lead are heavy metals that pose significant risks to plant health and therefore to human consumption.

The detection of a wider array of biochemicals in lettuce, spinach, and basil using spectral instruments can give a cheaper alternative to the traditional chemical analysis for advancing plant management, optimizing health, quality assessment, and nutrient optimization. As can be seen in the table below, all biochemical traits have good accuracies presented by the studies. Some of them, like potassium, phosphorus, and calcium, have many entries in the table, which shows the trend in research in recent years. Studies including biochemicals such as vitamin C, magnesium, and sulfur can give an edge to the grower in monitoring plant defense and enhancing nutritional value. Further studies should explore the potential of spectral imaging, focusing on the expansion of detectable biochemical traits.

5. Discussion

This systematic review gathered and showcased leading techniques in non-destructive estimation of biochemical traits in lettuce, spinach, and basil through spectral imaging technologies. The work shows advances in hyperspectral imaging, spectrometry, and multispectral imaging, with a focus on applications for accurate estimations of key chemical compounds. Those advanced techniques proved instrumental in achieving these results. Improvements in the field of deep learning help with improving regression and classification accuracy, which greatly reduce dependence on labeled data. Convolutional neural networks [68] are a very proficient tool to work with for extracting relevant features. The ability given to the researchers to analyze key wavelengths associated with biochemical traits has enabled precise quantification, which makes this approach feasible for researchers but also commercial agricultural applications. Spectral instruments are getting more accessible and cost-effective. With this technology getting cheaper, its adoption is likely to grow, which will give growers a valuable, reliable, and real-time tool that gives insight into plant well-being and can help with optimization of nutrient management and health monitoring. This indicates a potential gap in research that could be further explored because spectral imaging techniques inherently encounter several methodological and instrumental limitations. Key challenges include the unmixing of overlapping biochemical spectra, ensuring robust calibration and cross-validation across different instruments or imaging setups, and maintaining high repeatability of measurements. Intrinsic plant characteristics, such as leaf surface texture variations, leaf morphology, and phenological stages, further complicate accurate spectral measurements, potentially limiting broader applicability or requiring sophisticated algorithms for normalization and correction. These aspects warrant detailed attention in future research and systematic methodological guidelines to facilitate practical implementation. This review also showcases a wide use of machine learning techniques, which contributes to the good performance of spectral data analysis. Despite the impressive reported accuracies, notable limitations persist, primarily arising from the methodological heterogeneity across studies, unclear validation approaches, and potential biases introduced by language and publication restrictions. The focused scope on recent studies and specific plant species may overlook important insights available from broader contexts or earlier foundational research. Future research should focus on underrepresented plants like basil and refining estimation methods to achieve higher accuracies. Future research in the other biochemical traits should emphasize standardizing measurement and validation procedures, increasing methodological transparency, and developing more affordable spectral imaging devices and universally applicable algorithms to promote wider adoption and accessibility in agriculture. Due to the high accuracies reached, a real-life study should be conducted using spectral measurements to create a decision support system for biochemical management.