Applications of Machine Learning and Deep Learning for Foliar Nutritional Deficiency: A Systematic Review

Abstract

The automatic detection of foliar nutritional deficiencies through computer vision represents a promising alternative within precision agriculture practices, reducing dependence on laboratory analyses and the subjectivity associated with visual inspection. This systematic review maps and compares the application of machine learning (ML) and deep learning (DL) techniques to nutritional diagnosis across different crops, highlighting methodological trends, barriers to model adoption under field conditions, and existing research gaps. Following the PRISMA guidelines (PRISMA-P and PRISMA-2020), searches were conducted in the Scopus, IEEE Xplore, and Web of Science databases, using a defined time frame and explicit inclusion and exclusion criteria, resulting in 200 articles included (2012–2026; last search on 2 February 2026). The results indicate a predominance of DL-based approaches and RGB imagery, with applications concentrated in crops such as rice and in macronutrients, mainly nitrogen (N), phosphorus (P), and potassium (K), and report a marked increase in publications from 2020 onward. Although many studies report high performance, the evidence is largely derived from controlled environments and proprietary datasets, which limit model comparability, reproducibility, and generalization to real-world scenarios. Accordingly, the main research gaps include limited validation under field conditions, identified as the primary practical barrier; the underrepresentation of micronutrients and multiple-deficiency diagnosis; and the need for lightweight architectures suitable for deployment in mobile and edge-computing applications. It is concluded that ML and DL techniques offer promising alternatives for automated nutritional diagnosis; however, advances in data standardization, open-access datasets, and validation under real field conditions are essential for consolidating these technologies in practical applications.

1. Introduction

Adequate mineral nutrition is a decisive factor for crop productivity and vigor, and the delayed diagnosis and correction of nutritional deficiencies are responsible for substantial yield losses, ranging from 20% to 60% according to the literature [1,2,3]. Although laboratory analyses are considered reliable methods, the high cost per sample and the turnaround time for results may delay decision-making, leading to economic losses, particularly when fertilizer recommendations depend on sample collection, as shipping and processing may take several weeks [4,5,6].

In this context, diagnosis based on visual symptoms has been adopted as a complementary practice, since changes in leaf color, texture, and morphology often appear before yield reduction occurs. However, this approach relies on expert knowledge and is subject to subjective interpretation, especially when deficiencies of different nutrients exhibit similar symptoms or when environmental stress interferes with symptom expression [7].

In recent years, methods involving machine learning (ML) and deep learning (DL) models have advanced non-destructive diagnostic approaches by directly analyzing leaf images, driven by the increasing accessibility of libraries such as Scikit-learn, TensorFlow, and PyTorch 2.8, as well as the widespread availability of GPUs [8,9,10,11]. In this regard, studies conducted by Romualdo et al. (2014), Montes Condori et al. (2017), and Muthusamy and Ramu (2024) [12,13,14] demonstrated the application of convolutional neural networks (CNN) to diagnose nitrogen deficiency in maize, reducing the need for manual feature extraction, identifying complex visual patterns, and achieving high accuracy.

Despite this potential, significant gaps remain between reported performance in the literature and practical adoption, as most studies have been conducted under controlled conditions (laboratory or greenhouse). This poses challenges for real-world performance due to variability in illumination, background elements in images, and environmental stress factors affecting leaves [15,16]. Additionally, the limited availability of public, balanced, and standardized datasets with well-defined classes and visible symptoms hinders the classification of different nutritional deficiencies and their symptoms, while also compromising result reproducibility.

Although several reviews have addressed the application of artificial intelligence in agriculture, the literature still lacks a systematic analysis that comparatively examines image types and acquisition platforms, learning approaches and model architectures, and validation strategies, particularly contrasting controlled environments with field conditions. Such analysis is essential to highlight how practical barriers challenge generalization, reproducibility, and operational feasibility, ultimately transforming experimental results into deployable tools.

Within this framework, the present systematic review compiles, organizes, and analyzes studies that applied ML and DL models for the identification, diagnosis, and classification of nutritional deficiencies in agricultural crops using image-based approaches, following PRISMA guidelines (in the Supplementary Materials). The review aims to map methodological and technological trends, identify research gaps—particularly those related to field validation and data availability—and guide future studies toward practical adoption. The guiding research question of this review is: “How have computer vision and ML techniques been applied to identify and classify nutritional deficiencies in plants?”

2. Materials and Methods

2.1. Type of Study and PICo Strategy

This study is grounded in the guidelines of the PRISMA-P 2015/2020 method [17,18], which provides the complete framework required to conduct a rigorous systematic review consistent with the academic standards expected in the fields of agriculture and artificial intelligence, as detailed in

Table 1. Stages of the systematic review according to the PRISMA-P 2015/2020 method.

#	Step	Description
1	Title	Specification of the topic and indication that it is a systematic review protocol
2	Registration	Indicate whether the protocol was registered in PROSPERO; however, reviews in the agricultural field do not require registration
3	Authors	Provide all relevant information about the authors, affiliations, and contact details
4	Funding Support	Indicate any financial or institutional support
5	Conflicts of Interest	Report potential conflicts of interest
6	Rationale	Explain the relevance of the topic and the existing research gaps
7	Objectives	State the objectives of the review along with the review question(s), using the PICO format, which may be adapted to the context
8	Eligibility Criteria	Define the inclusion and exclusion criteria for the studies
9	Information Sources	List all databases to be consulted
10	Search Strategy	Specify the search terms and strategies, including keywords and Boolean operators
11	Record Management	Describe how records are managed and organized
12	Study Selection	Explain how studies will be selected
13	Data Extraction	Detail the data extraction process
14	Data Items	Define the data items to be extracted according to the study scope
15	Outcomes/Prioritization	Specify the prioritization of studies and outcomes according to the defined metrics
16	Risk of Bias	Describe the analyses conducted to identify potential risks of bias
17	Data Synthesis	Explain how the data will be analyzed, such as through meta-analysis or descriptive analysis

Moher et al. (2015); Page et al. (2021) [17,18].

.

Accordingly, this systematic review focuses on the application of ML and DL techniques for the detection and classification of foliar nutritional deficiencies in agricultural crops. Thus, the guiding research question, formulated according to the PICo framework (Population, Interest, and Context), seeks to answer: “How have computer vision and ML techniques been applied to identify and classify nutritional deficiencies in plants?” The PICo strategy was structured as follows: (P) plants/agricultural crops; (I) computational techniques based on computer vision (ML and DL); and (Co) identification/detection/classification of nutritional deficiencies in leaves.

2.2. Data Sources and Search Strategy

The searches were conducted in three major databases widely used and recognized for scientific literature retrieval: Scopus, Web of Science (WoS), and IEEE Xplore. Google Scholar was not selected as a primary database due to its limitations regarding filtering capabilities and duplicate record retrieval. Searches were carried out in two rounds: the initial search on 2 December 2025 and the final update on 2 February 2026, including articles published in 2026. The time span covered by the search ranged from 2012 to 2026.

Accordingly, to obtain a body of literature that met the key requirements of the PRISMA-P (2015) protocol and the PRISMA 2020 reporting guidelines, the initial search filters within the databases were defined as follows:

• Document type: “Article” and “Conference paper”—conference papers were included due to their relevance in presenting applied studies and methodological innovations at scientific events, such as conferences, which is a characteristic of the artificial intelligence field.
• Source type: “Journal” and “Conference proceeding”—corresponding to the selected document types.
• Language: English—to ensure linguistic consistency in the analyses and because it is the predominant language in scientific publications within the field.

Distinct search strategies were developed using Boolean operators, in accordance with the indexing specifications of each database, to retrieve high-quality articles and applications aligned with the research question.

Table 2. Search terms and Boolean operators used in each database.

Database	Search Terms and Boolean Operators
Scopus	(TITLE-ABS-KEY(“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network” OR “CNN” OR “computer vision” OR “image processing” OR “pattern recognition”)) AND (TITLE-ABS-KEY(“nutrient deficiency” OR “nutritional deficiency” OR “nutrient diagnosis” OR “foliar diagnosis” OR “leaf diagnosis” OR “deficiency detection” OR “deficiency identification” OR “nutrient stress” OR “nitrogen deficiency” OR “phosphorus deficiency” OR “potassium deficiency”)) AND (TITLE-ABS-KEY(“crop” OR “cereal” OR “grain” OR “sorghum” OR “maize” OR “corn” OR “soybean” OR “soy” OR “rice” OR “wheat” OR “cotton” OR “tomato” OR “plant” OR “leaf” OR “agriculture” OR “agricultural”))
Web of Science (WoS)	TS = (“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network” OR “CNN” OR “computer vision” OR “image processing” OR “pattern recognition”) AND TS = (“nutrient deficiency” OR “nutritional deficiency” OR “nutrient diagnosis” OR “foliar diagnosis” OR “leaf diagnosis” OR “deficiency detection” OR “deficiency identification” OR “nutrient stress”) AND TS = (“crop” OR “cereal” OR “grain” OR “sorghum” OR “maize” OR “corn” OR “soybean” OR “soy” OR “rice” OR “wheat” OR “cotton” OR “tomato” OR “plant” OR “leaf” OR “agriculture”))
IEEE Xplore	((“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network” OR “CNN” OR “computer vision” OR “image processing”) AND (“nutrient deficiency” OR “nutritional deficiency” OR “nutrient diagnosis” OR “foliar diagnosis” OR “deficiency detection”) AND (“crop” OR “cereal” OR “grain” OR “sorghum” OR “maize” OR “corn” OR “soybean” OR “rice” OR “wheat” OR “cotton” OR “tomato” OR “plant” OR “agriculture”))

presents the search terms employed for each database.

2.3. Eligibility Criteria

After the database search, eligibility criteria were established, followed by exclusion criteria, to ensure that only articles aligned with the research question and providing relevant and reliable analyses were included in this review. The eligibility criteria served as a clear framework to maintain focus on the research topic, allowing studies that did not meet these criteria to be excluded from the analysis. Accordingly, the eligibility and exclusion criteria are presented in the following topics, as summarized in

Table 3. Inclusion and exclusion criteria.

Inclusion Criteria	Exclusion Criteria
(i) applied ML and/or DL to images to identify, detect, and classify foliar nutritional deficiencies;	(i) The analysis of the abstracts, along with the conclusion, excluded articles that were out of scope;
(ii) analyzed plants/agricultural crops;	(ii) Article without access;
(iii) presented methodology and results (e.g., metrics, accuracy, F1-score, when applicable);	(iii) Some abstracts lack sufficient methodological data;
(iv) were published within the period from 2012 to 2026;	(iv) Studies outside the scope (e.g., disease/pest detection without a nutritional focus or nutritional studies without the use of imaging and ML/DL) and after verifying the methodology).
(v) were available in full-text format.

.

2.4. Article Selection Process

The article selection process followed five stages, as recommended by the protocol, ensuring methodological transparency and reproducibility. First, the reviewers were defined, and the process was conducted by two independent reviewers (the main author and J.R.R.), strictly following the inclusion and exclusion criteria presented in Table 3. Disagreements were resolved through discussion and consensus, and when necessary, a third reviewer was consulted to minimize bias.

After defining the total number of records retrieved from the three databases (Scopus = 790; WoS = 289; IEEE = 262), the data were exported and compiled into a Microsoft Office Excel spreadsheet. Duplicate records were identified by comparing titles, authors, publication year, and Digital Object Identifier (DOI). Duplicate removal followed specific criteria: articles with identical DOIs were considered duplicates based on analysis performed using Mendeley; for articles without a DOI, titles and authors were manually checked. When the same article appeared in multiple databases, priority was given to the database providing the most complete metadata, following the order Scopus > WoS > IEEE. At the end of this process, 546 duplicate records were removed, resulting in 795 unique articles for the next phase, namely screening.

The titles of the 795 articles were individually evaluated to assess adherence to the eligibility criteria (Table 3). Excluded articles clearly exhibited exclusion criteria in their titles, such as exclusive focus on diseases or pests, applications in non-agricultural domains, review studies without methodological contribution, and studies that did not involve computer vision or ML approaches.

In addition, a specific justification for the exclusion of each article was recorded, ensuring full traceability of the process. At the end of this stage, 451 articles were excluded, leaving 344 articles for the abstract screening phase.

In the fourth stage, the remaining articles were analyzed in detail by jointly reviewing titles and abstracts to verify compliance with the eligibility criteria (Table 3), along with documenting specific reasons for exclusion. Even when some articles did not provide sufficient methodological details at this stage, they were retained for the subsequent phase to allow full-text analysis, enabling more rigorous decision-making and minimizing inappropriate exclusions.

The main reasons for exclusion were similar to those identified during the title screening phase, with the addition of studies focused on phenotyping, yield prediction, and topics unrelated to nutritional diagnosis, as well as studies that did not independently categorize nutritional deficiencies. At this stage, 37 articles were excluded, resulting in 307 articles eligible for full-text review.

The final stage consisted of full-text reading to verify compliance with the eligibility criteria. The SciSpace tool (https://scispace.com/) was used to support systematic analysis and the extraction of relevant information from the articles, enabling automated analysis of specific sections, export of structured tables, and interactive querying of article content to clarify methodological aspects and facilitate comparisons. This process contributed to a more consistent selection aligned with the research topic and supported the exclusion of studies with inadequate methodology and/or insufficient information.

At the end of this stage, 107 articles were excluded for reasons including: inaccessible articles or lack of full-text availability, even though the CAPES Journal Portal, despite being eligible based on title and abstract, which prevented detailed analysis; inadequate methodology, such as absence of ML or DL application, lack of image-based approaches, insufficient experimental validation, or inappropriate scope, also described in Table 3.

To maintain methodological rigor, a detailed record of the entire selection process was maintained in an Excel spreadsheet, including essential metadata, ensuring transparency and reproducibility of this systematic review. After completing all stages, 200 articles were included in the review for data synthesis and detailed analysis of their methodologies and applications. The complete PRISMA flow diagram of the process, including the number of articles at each stage, is presented in Figure 1, in accordance with the steps required by the PRISMA-P 2015 and PRISMA 2020 guidelines.

It should be noted that duplicate articles and those without access to the full text were excluded from the article database and recorded with their respective reasons for exclusion in an auxiliary spreadsheet during the selection process. Likewise, studies that analyzed diseases, pests, and nutritional deficiencies without distinguishing them into independent classes were also excluded, as they did not meet the eligibility criterion requiring nutritional deficiencies to be treated as a specific category. Thus, only studies employing appropriate methods and presenting a focus aligned with the objectives of this review were included

2.5. Data Extraction, Organization, and Synthesis

The data extraction process was conducted in a systematic and structured manner for each of the 200 articles, with information recorded according to predefined categories, facilitating comparative analyses and ensuring that relevant aspects were captured uniformly. For each article, bibliometric information was recorded, including year of publication, authors, indexing databases, publishing journal or conference, and number of citations. Regarding study characteristics, information was collected on the crop(s) analyzed, the research objective, the nutrients and corresponding deficiencies investigated—specifying whether they were analyzed in conjunction with diseases and/or pests—the image acquisition tools employed (e.g., sensors, drones, cameras, mobile devices, satellites), the acquisition context (laboratory, greenhouse, or field), and dataset characteristics, such as number of classes, balancing strategies, use of data augmentation techniques, and sample size.

Within the methodological approach, information was documented on the architectures employed, including the ML and/or DL models used and whether they were applied jointly; feature extraction strategies, distinguishing between manual descriptors (e.g., texture analysis) and automatic feature learning; validation methods, such as train–validation–test splits or k-fold cross-validation; performance metrics; and the main results achieved by the models. Finally, information on study limitations and perspectives was recorded, including suggestions for future research and potential application under real-world conditions.

Microsoft Office Excel spreadsheets were used to standardize information from the three databases into a single dataset, maintaining organized records and enabling descriptive, comparative, and exploratory analyses. The data were structured into categorized columns to facilitate information processing and visualization, as illustrated in Figure 2.

This structure enabled the analysis of frequencies, rankings, and the application of dynamic filters, allowing the identification of patterns and trends across the studies.

2.6. Descriptive Data Synthesis

With the data organized, descriptive analyses of the studies were conducted, examining the methodologies employed across the articles, their similarities and differences, and the main findings. The analysis began with temporal assessments to identify which architecture achieved the best performance, the primary tools used to capture input data (images), and the most frequently studied nutritional deficiencies. In addition, the analyses evaluated which models demonstrated generalization capability and whether they were applied under real-world (field) conditions.

To identify growth or decline in subareas of computer vision and ML applied to nutritional diagnosis, temporal analyses were performed, including the distribution of publications between 2012 and 2025. To complement these analyses, publications were mapped according to countries and institutions, indicating major research centers in this field and international collaboration networks.

The analysis also examined the most studied crops, the nutrients most frequently investigated, and the main research challenges. Another aspect evaluated was the type of sensors and platforms used, including comparisons between studies employing RGB images versus hyperspectral imagery, as well as those conducted in controlled environments (laboratory or greenhouse) versus field applications.

Regarding computational methods, the analysis identified the most used architectures and compared the performance of classical methods (SVM, Random Forest, KNN) with deep neural networks (CNN, ResNet, VGG, among others). Based on this comparison, reported accuracy ranges were identified to assess model performance and the combinations of data acquisition tools and models that yielded the best results. The importance of additional performance metrics was also highlighted, examining consistency among them, such as recall (sensitivity), given that false negatives tend to have a greater impact on practical applications than false positives.

Finally, the descriptive synthesis identified nutritional deficiencies that remain underexplored despite their agronomic importance, as well as key issues highlighted in the literature and economically important crops that have been underrepresented. Methodological challenges were also identified, including the risk of overfitting in small datasets, lack of field validation, and limited model generalization across crops, nutrients, and environmental conditions.

2.7. Bibliometric Network Analysis

For the bibliometric analyses, Python (version 3.14) scripts were employed using the Google Colab (Colaboratory) platform. The analyses were conducted based on the metadata of the 200 articles selected for this systematic review, which were organized in a standardized spreadsheet containing information on title, authors, keywords, abstract, year of publication, and institutional affiliation, ensuring methodological consistency across all stages of the research.

To enable data processing by the software, the spreadsheet was exported in .CSV or Excel format. Accordingly, the analyses performed to generate the results of this review included keyword co-occurrence techniques, which identified the most frequently used terms and their relationships, with a minimum threshold of two to three occurrences to ensure statistical relevance.

Co-authorship analyses were conducted to map collaborations among authors, institutions, and countries, in addition to temporal analyses that assessed the evolution of research over time according to eligibility criteria based on year of publication. Finally, clustering graphs were generated to detect related topics by grouping them according to the modularity-based method.

The results generated from the bibliometric analyses supported the discussion of descriptive and qualitative findings by highlighting research gaps, the most commonly used ML and DL methods, and the growth of data acquisition technologies and image types, such as UAV and hyperspectral imagery, as well as advances in early detection of nutritional deficiencies, providing an overview of the evolution of these technologies in agricultural applications.

2.8. Risk of Bias and Reproducibility Assessment

Considering the diversity of datasets and the different methodological applications, an operational assessment of risk of bias and reproducibility was conducted, supported by the results and indicators extracted from the articles. Accordingly, the following items were considered (yes/no/partial, when applicable):

• Type of validation (field versus controlled environment).
• Description of datasets and sample size.
• Clarity in dataset partitioning, including training/testing splits and prevention of data leakage.
• Completeness of reported performance metrics.
• Public availability of data and/or code.
• Description of the experimental protocol and essential hyperparameters.

To contextualize the limitations reported by the studies, model performance, potential overestimation, and challenges related to generalization under real-world scenarios, these indicators were discussed in an integrated manner throughout the results synthesis.

2.9. Final Synthesis

The combined descriptive and bibliometric analyses made it possible to map the state of the art, identify methodological gaps, and outline directions for future research, in addition to answering the guiding question of this review. The consolidated results of these analyses are presented in Section 3 and Section 4 (Results and Discussion).

The complete dataset comprising the 200 articles included in this review, including bibliometric information, technical characteristics, and the main performance metrics, is available in the OSF repository under DOI 10.17605/OSF.IO/DUZH8. The script used in the analyses, downloaded from Google Colab under the file extension, is attached to the OSF repository too.

3. Results

For the presentation and discussion of the results, the section is divided into two interconnected parts that provide promising insights into the topic under investigation. The first part refers to the results of the bibliometric analysis of the 200 articles [12,13,14,15,16,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200,201,202,203,204,205,206,207,208,209,210,211,212,213] derived from the PRISMA-P protocol extraction process and selected from the main scientific databases. Accordingly, the data are presented in the form of graphs, tables, and figures to enhance visualization and summarize the main findings, providing an overall perspective of the field. The second part focuses on the in-depth analysis of the most relevant articles, defined by their higher number of citations and methodological contributions, offering a detailed examination of the main technologies employed and their evolution within agricultural applications.

3.1. Bibliometric Analysis

The 200 articles are distributed between 2012 and 2026, as shown in Figure 3, illustrating their distribution over time. Initially, only one study was published in 2012, followed by a period of low publication frequency up to 2019. From 2020 onward, a continuous growth in research output is observed, with 12 articles published in 2020. In 2024 and 2025, a total of 97 articles were published, accounting for 48.50% of the total.

In terms of geographic distribution, as illustrated in Figure 4, publications were concentrated in India, which accounted for 87 publications (43.50% of the total). This was followed by China with 21 publications (10.5%), Indonesia with 10 publications (5.0%), Malaysia also with 10 publications (5.0%), and the Philippines with 8 publications (4.0%). Brazil, the United States, Peru, Bangladesh, and Sri Lanka each contributed five articles (2.5%).

Regarding indexing sources, most articles were retrieved from Scopus, with 117 articles (58.5%), followed by IEEE Xplore with 61 articles (30.5%) and Web of Science (WoS) with 22 articles (11%). After consolidating records from the three databases, duplicate articles across databases were removed prior to the final screening.

The publication venues, including journals and conference proceedings, are summarized in

Table 4. Main publication venues—Top 10.

Journal	Number of the Articles	% Total (N = 200 Articles)	Number of Citations (Total)
IEEE	64	32.0	226
Elsevier BV	12	6.0	143
MDPI	10	5.0	251
Springer	9	4.5	29
Agronomy	5	2.5	34
Computers and Electronics in Agriculture	4	2.0	168
Nature Research	3	1.5	12
Frontiers in Plant Science	2	2.0	40
Remote Sensing	2	2.0	87
Electronics (Switzerland)	2	2.0	79

. IEEE Access accounted for 64 articles (32%) with a total of 226 citations, followed by Elsevier BV with 12 articles (6%) and 143 total citations, and MDPI with 10 articles (5%) and 251 total citations. Computers and Electronics in Agriculture stands out in terms of citation count, with a total of 168 citations.

Regarding publication type, 116 studies (58%) were journal articles and 84 were conference papers (42%). In addition, 2021 was the year with the highest number of citations, totaling 635 citations (22.01% of the total, N = 2885). Notably, 2018 represents a point of interest, as five published articles accumulated 504 citations (17.47%).

Among the publication authors, it was observed that the most productive authors published only two articles each within the analyzed period. The most cited studies and their main conclusions are presented in

Table 5. Research details of the 10 most cited authors.

Reference	Database	Article	Authors	Conclusion	DOI/Year/Citations
[196]	Scopus	An explainable deep machine vision framework for plant stress phenotyping	Ghosal et al.	Proposes an explainable framework that identifies, classifies, and quantifies foliar stresses (including nutritional deficiencies) from images, demonstrating robustness and potential for large-scale deployment.	https://doi.org/10.1073/PNAS.1716999115 (2018)—454
[205]	Scopus	A comparative study of deep CNN in forecasting and classifying the macronutrient deficiencies on development of tomato plant	Tran et al.	Demonstrates that CNN models (Inception-ResNet v2 and Autoencoder) and ensemble approaches improve the prediction and classification of deficiencies (Ca, K, and N) in tomato plants, with the ensemble achieving superior validation performance.	https://doi.org/10.3390/app9081601 (2019)—124
[184]	Scopus	A deep learning approach to measure stress level in plants due to Nitrogen deficiency	Azimi, S et al.	It proposes a 23-layer CNN that outperforms traditional methods in classifying nitrogen stress levels using images of shoots.	https://doi.org/10.1016/j.measurement.2020.108650 (2021)—109
[62]	WoS	Estimation of paddy rice nitrogen content and accumulation both at leaf and plant levels from UAV hyperspectral imagery	Wang, L et al.	It demonstrates that the use of UAV-mounted hyperspectral sensors enables high-accuracy estimation of nitrogen concentration and accumulation at different plant levels.	https://doi.org/10.3390/rs13152956 (2021)—86
[42]	Scopus	Nutrients deficiency diagnosis of rice crop by weighted average ensemble learning	Talukder & Sarkar.	Shows that a weighted ensemble of pre-trained CNN enhances performance and achieves high test accuracy for diagnosing nutritional deficiencies in rice crops.	https://doi.org/10.1016/j.atech.2022.100155 (2023)—80
[26]	Scopus	Using deep convolutional neural networks for image-based diagnosis of nutrient deficiencies in rice	Xu, Z et al.	It validates that deep neural net-works can automatically diagnose ten types of deficiencies in rice with performance superior to human visual inspection. The best model obtained test precision of 97,44% identifying N, P and K deficiencies.	https://doi.org/10.1155/2020/7307252 (2020)—79
[142]	Scopus	Comparative performance of four CNN-based deep learning variants in detecting Hispa pest, two fungal diseases, and NPK deficiency symptoms of rice (Oryza sativa)	Dey et al.	Compares CNN models and shows that performance depends on the stressor and dataset (public and field), with specific models outperforming others for different deficiencies and diseases.	https://doi.org/10.1016/j.compag.2022.107340 (2022)
[202]	Scopus	Airborne hyperspectral imaging of nitrogen deficiency on crop traits and yield of maize by machine learning and radiative transfer modeling.	Wang, S et al.	It concludes that integrating radiative transfer models with machine learning improves the prediction of biomass and yield in maize under nitrogen stress.	https://doi.org/10.1016/j.jag.2021.102617 (2021)—72
[65]	WoS	Ensemble Averaging of Transfer Learning Models for Identification of Nutritional Deficiency in Rice Plant	Sharma, M et al.	It demonstrates that combining multiple pre-trained models in an ensemble reduces individual classification errors and enhances robustness in nutritional diagnosis.	https://doi.org/10.3390/electronics11010148 (2022)—71
[25]	WoS	Using Deep Convolutional Neural Network for Image-Based Diagnosis of Nutrient Deficiencies in Plants Grown in Aquaponics	Taha, M. F. et al.	It demonstrates the technical feasibility of DL for monitoring vegetable health in aquaponic systems, enabling rapid and non-invasive diagnostics.	https://doi.org/10.3390/chemosensors10020045 (2022)—70
[137]	Scopus	Deep Learning Based Disease, Pest Pattern and Nutritional Deficiency Detection System for “Zingiberaceae” Crop	Waheed et al.	Evaluates DL on a real ginger dataset and reports strong performance in classifying diseases, pests, and nutritional deficiencies, highlighting potential for real-time applications.	https://doi.org/10.3390/agriculture12060742 (2022)

Note: Citation counts were extracted from the Scopus, IEEE e WoS database on 2 February 2026 and refer to the number of citations reported by that source at the time of data extraction.

, along with information on the title, authors, and DOI/Year/Citations.

3.2. Agronomic Characteristics of the Studies

The distribution of crops addressed in the 200 articles is presented in

Table 6. Distribution of studied crops—Top 15.

Crop	Frequency	Percentage (%)	Predominant Region in the Articles	Technical Remarks
Rice	44	22.00	Asia	Crop with the largest data volume, mainly focused on NPK deficiencies
Multi-crop studies	33	16.50	Asia	Emphasis on model generalization for universal mobile applications
Maize	15	7.50	Asia and Americas	Brazil stands out, especially studies conducted at ESALQ/FZEA
Tomato	12	6.00	Asia	Focus on controlled environments (greenhouses) and bench-scale systems
Coffee	10	5.00	Latin America and Asia	Brazil leads technical publications on AI applications for coffee
Banana	8	4.00	Asia	Many studies specifically address potassium deficiency
Pepper	6	3.00	Asia/S. America	Focuses on multi-class classification of NPK and micro-element stressors like Magnesium.
Oil palm	6	3.00	Asia	Frequently utilizes SVM and RBF networks for detection under uncontrolled daylight conditions.
Lettuce	5	2.50	Asia	Emphasis on hydroponic systems and plant factories
Peanut	4	2.00	Asia	Studies focused on foliar diagnosis of nitrogen and chlorophyll via smartphones
Citrus	4	2.00	Asia/Europe	Often distinguishes between nutritional chlorosis and Huanglongbing (greening) disease symptoms.
Wheat	4	2.00	Asia/Europe	High application of UAV-based RGB and multispectral imagery for field-scale nitrogen monitoring.
Bean	3	1.50	S. America/Asia	Studies emphasize early-stage detection of Nitrogen and Phosphorus through texture and color descriptors.
Cucumber	2	1.00	Asia	Extensive use of hyperspectral images and regression models for nitrogen estimation
Cotton	2	1.00	Asia	Growing focus on UAV/drone-based detection
Other crops (N = 28)	24	12.00	Variable	Represents the highest diversity of tropical and horticultural crops

Note: The category “multi-crop” refers to studies that evaluated two or more crops within the same experiment.

. Rice was the predominant crop among the studies, accounting for 44 articles (22%), while studies considering multiple crops comprised 33 articles (16.5%). This was followed by maize, investigated in 15 studies (7.5%), and tomato, with 12 studies (6.0%). Other crops appeared at lower frequencies, representing a diverse set of 28 crops grouped under the category “Others,” totaling 24 articles (12%).

A total of 672 nutrient mentions were recorded across the 200 articles, corresponding to an average of 3.36 nutrients per article and a median of 3.0. Nutrient deficiencies are presented in

Table 7. Distribution of studied nutritional deficiencies (N = 672 mentions).

Category	Nutrient	Mentions	% Total	% Category	Characteristic Visual Symptom
Primary macronutrients	N	153	22.80	40.20	Progressive generalized chlorosis
	K	126	18.80	33.10	Marginal necrosis of older leaves
	P	102	15.20	26.80	Purplish coloration, reduced growth
	Subtotal	381	56.7%	100%
Secondary Macronutrients	Ca	87	12.90	57.6	Apical necrosis (blossom-end rot)
	Mg	47	7.00	31.10	Interveinal chlorosis of older leaves
	S	17	2.50	11.30	Uniform chlorosis of young leaves
	Subtotal	151	22.50%	100%
Micronutrients and Other elements	Fe	77	11.50	55.00	Interveinal chlorosis of young leaves
	Manganese (Mn)	20	3.00	14.30	Necrotic spots, chlorosis
	Zn	19	2.80	13.60	Small leaves, shortened internodes
	B	15	2.20	10.70	Deformation of new organs
	Cu	3	0.40	2.10	Shoot tip wilting
	Mo	2	0.30	1.40	Chlorosis with necrosis
	Other elements *	4	0.6	2.9	Variable
	Subtotal	140	20.80%	100%
OVERALL TOTAL		672	100%

* Silicon (Si), Nickel (Ni), Sodium (Na), and Chlorine (Cl)—one each. Note: Percentages were calculated based on a total of 672 individual nutrient mentions across 200 studies. The mean number of nutrients per study was 3.4 (median: 3.0). A single study may investigate multiple nutrients; therefore, the total number of mentions exceeds the number of articles.

, in which percentages were calculated based on the total number of mentions (n = 672). Overall, macronutrients accounted for 79.2% of the mentions, whereas micronutrients represented 20.8%.

Among the primary macronutrients, nitrogen (N) was the most frequently studied, with 153 mentions (22.8% of the total), followed by potassium (K) with 126 mentions (18.8%) and phosphorus (P) with 102 mentions (15.2%). Secondary macronutrients accounted for 151 mentions (22.5%), with emphasis on calcium (Ca), with 87 mentions (12.9%), magnesium (Mg) with 47 mentions (7.0%), and sulfur (S) with 17 mentions (2.5%).

Among the micronutrients, iron (Fe) was the most frequently investigated, with 77 mentions (11.5% of the overall total and 55% of micronutrient mentions), followed by manganese (Mn) with 20 mentions (3.0%), zinc (Zn) with 19 mentions (2.8%), and boron (B) with 15 mentions (2.2%). Nutrients with low frequency included copper (Cu) with three mentions (0.4% of the total) and molybdenum (Mo) with two mentions (0.3%), as well as other micronutrients with only one mention each (silicon—Si; nickel—Ni; sodium—Na; and chlorine—Cl).

Another approach analyzed was the ability to identify multiple nutrient deficiencies simultaneously, as illustrated in Figure 5. The results indicate a low incidence of studies addressing six or more nutrients, totaling 13 articles (6.5%).

Studies that addressed analyses involving more than six nutrients—characterized as multi-nutrient studies—were infrequent, comprising only 25 articles (12.5%). Thus, only one out of every eight studies demonstrated the capability to diagnose six or more deficiencies simultaneously.

3.3. Models and Architecture

The analysis of category distribution (n = 200 articles) indicates the predominance of DL models, which were present in 122 articles (61%), followed by ML models in 57 articles (28.50%) and hybrid approaches (ML + DL) in 21 articles (10.50%). In addition to the article-level counts, the number of mentions of each category across the analyzed studies was also recorded, corresponding to the occurrences registered during the information extraction process. DL was mentioned 403 times, ML 152 times, and hybrid ML + DL approaches 55 times, as shown in Figure 6.

In the studies that reported accuracy values (n = 179), DL models achieved a mean accuracy of 94.2%, with a median of 96.42% [CS54.1]. The interquartile range indicates that 50% of the studies reported accuracy between 90% and 98%. In studies using classical ML approaches, the mean accuracy was 91.11%, with a median of 92.30%. Outliers were observed across all categories, with performance below 80%; overall, the upper limit across all categories reached 100% accuracy.

Regarding methodological approaches, as illustrated in Figure 7, 99 studies (49.5%) conducted comparative analyses, followed by 66 studies (33%) that compared their results with baseline models, and 20 articles (10%) that employed transfer learning. In addition, 76 articles (38%) proposed the adoption of custom models or architectures developed by the authors.

Detailing the models and architectures, as presented in

Table 8. Most frequently used architectures and models—Top 10.

Model	Mentions	%	Category	Characteristics
VGG-16	70	11.48	DL	Sequential architecture
ResNet	65	10.66	DL	Use of skip connections
Random Forest	57	9.34	ML	Decision tree ensemble
SVM	52	8.52	ML	Kernel-based classifier (Support Vector Machine)
Custom (non-standardized) CNNs	37	6.07	DL	Study-specific architectures developed for individual applications
Inception	27	4.43	DL	Use of multi-scale convolution modules
KNN	23	3.77	ML	A simple, non-parametric instance-based learning algorithm that classifies based on feature proximity.
MobileNet	20	3.28	DL	Focus on computational efficiency and mobile deployment
ANN	20	3.28	DL	Artificial Neural Networks
AlexNet	13	2.13	DL	A pioneering deep CNN featuring stacked convolutional layers, ReLU activation, and dropout for regularization.
Other architectures	226	37.07	Mixed	Includes 108 additional models (AlexNet, YOLO, ViT, KNN, etc.)

, a total of 610 mentions were recorded across the 200 articles. It was observed that 141 articles investigated only a single model, resulting in an average of 3.05 models per article. The most frequently used architectures were VGG16, with 70 mentions (11.5% of the total), ResNet50 with 65 mentions (10.7%), Random Forest with 57 mentions (9.3%), SVM with 52 mentions (8.5%), and CNN with 37 mentions (6.1%). Other relevant architectures included Inception with 27 mentions (4.43%), KNN with 23 mentions (3.77%), MobileNet with 20 mentions (3.28%), ANN with 20 mentions (3.28%), and AlexNet with 13 mentions (2.1%). In addition, other models categorized as “Others,” such as AlexNet, Decision Tree, YOLO, ViT, DenseNet, among others, accounted for 226 mentions (37.1%), corresponding to a total of 108 distinct models.

3.4. Image Modalities and Acquisition Platforms

Most studies used custom datasets, accounting for 156 articles (78%). Dataset sizes ranged from 20 to 100,000 images, with a mean of 3254 images and a median of 1440 images. Figure 8 illustrates the distribution of studies according to dataset size ranges.

Regarding the use of public datasets, 25 articles (12.50%) were identified, with PlantVillage being the most frequently used dataset. Other public datasets cited included RoCoLe, Kaggle, and CIMMYT-Wheat. In 12 studies (6%), a mixed approach was adopted, combining public datasets with proprietary data. Data availability was reported in only 30 studies (15%), whereas 170 studies (85%) did not share their datasets. Furthermore, larger datasets exhibited higher data-sharing frequencies, with availability rates of 10.1% for small datasets, 13.4% for medium-sized datasets, and 28.2% for large datasets.

Regarding class distribution, 86 articles (43%) reported dataset imbalance. As a mitigation strategy, 39 studies (19.5%) applied techniques such as data augmentation. The characteristics of image types and acquisition platforms are detailed in

Table 9. Image types and acquisition methods.

Category	Type	Frequency	Percentage (%)	Typical Resolution	Characteristics
Image resolution	RGB	172	86.00	400–700 nm (visible)	Low cost, high availability
	Hyperspectral	14	7.00	100–400 bands	High spectral resolution
	Multispectral	4	2.00	4–12 bands + NIR	Includes near-infrared
	Multimodal (≥2 types)	10	5.00	Various	Various
Acquisition platform	Digital Camera	136	68.00	<0.1 cm/pixel	High spatial resolution
	UAV/Drone	17	8.50	0.5–5 cm/pixel	Medium-area coverage
	Smartphone	40	29.4%	<0.5 cm/pixel	Accessible and portable
	Other platforms	31	15.50	Various	Various

, including their frequency and typical attributes.

A predominance of RGB imagery was observed, being used in 172 articles (86% of the total). To a lesser extent, hyperspectral images were employed in 14 studies (7.0%), followed by multispectral images in four studies (2.0%). Studies that adopted more than one image modality (multimodality, e.g., RGB combined with another sensor) accounted for 10 studies (5%).

Regarding acquisition platforms, digital cameras were employed in 136 studies (68% of the total), followed by UAVs/drones in 17 studies (8.5%) and smartphones in 16 articles (8%). Other types of platforms, such as satellites equipped with Sentinel-2 and Planet sensors, as well as scanners, robots, and related systems, accounted for 31 studies (15.5%).

3.5. Validation and Experimental Conditions

The analysis of model performance, based on 179 articles (89.5% of the total) that reported accuracy values, indicates a mean accuracy of 9.3%, a median of 95%, and a standard deviation of 6.89%, as shown in Figure 9.

The boxplot (Figure 9) shows a concentration of accuracy values in the upper quartile, with a median of 95%. The interquartile range (box), spanning from 89.05% to 97.89%, indicates a high concentration of results at elevated accuracy levels, although the presence of outliers highlights isolated challenges in lower-performance scenarios.

The distribution of results showed that 145 articles (81%) achieved accuracies higher than 90%, of which 51 articles (28.5%) reported accuracies in the range of 90–95%, 71 articles (39.7%) reported accuracies between 95% and 99%, and 23 studies (12.80%) achieved accuracies above 99%. Accuracies below 90% were reported in a total of 34 studies (19%). Other reported metrics included precision, reported in 115 studies, with a mean of 94.57%, a median of 95.90%, and a standard deviation of 4.56%; the F1-score, reported in 107 studies, with a mean of 93.80%, a median of 95%, and a standard deviation of 6%; and recall, reported in 114 studies, with a mean of 94.48%, a median of 96%, and a standard deviation of 4.72%.

Additional metrics reported included AUC in eight articles (4%), with a mean value of 0.97; mAP@IoU in three detection studies (1.5%), with a mean of 54.73%; and, in regression studies, R² reported in 18 articles (9%), with a mean of 0.84, RMSE reported in 14 studies (7%), with a mean of 51.31%, and MAE reported in eight articles (4%), with a mean of 6.41%.

Comparing performance across model types showed a mean accuracy of 94.20%, a median of 96.42%, and a standard deviation of 6.37% for DL models (N = 115), whereas classical ML models (N = 45) achieved a mean accuracy of 91.11%, with a median of 92.30% and a standard deviation of 7.11%. Hybrid approaches (N = 19) resulted in a mean performance of 92.74%, a median of 95.63%, and a standard deviation of 8.49%, as shown in Figure 8.

Among DL architectures, ResNet-50 (N = 17) exhibited a mean performance of 93.84%, followed by custom CNN (N = 11) with 91.95%, VGG16 (N = 7) with 89.16%, and MobileNetV2 (N = 5) with 9097%. Among ML architectures, Random Forest (N = 20) achieved a mean accuracy of 90.97%, ANN (N = 6) achieved 92.83%, and SVM achieved 85.88% [CS60.1].

Finally, regarding methodological practices, 56 articles (28%) employed cross-validation, whereas the majority (N = 144) used a simple train–test split, representing 72% of the studies. Comparisons with baseline models were conducted in 190 studies (95%), and inference time was reported throughout the process, with a median of 107.50 ms.

4. Discussion

The literature demonstrates a substantial increase in the use of computer vision combined with ML and DL for the diagnosis of foliar nutritional deficiencies over the analyzed period (2012–2025), particularly from 2020 onward, when a strong expansion in publications was observed. This trend can be explained by the maturation of the applied AI ecosystem, including increased computational power, the consolidation of libraries and frameworks such as Scikit-learn, TensorFlow, and PyTorch 2.8, as well as improved access to development and deployment platforms, including cloud services and mobile devices.

In the agronomic context, the primary motivation lies in reducing costs and response time associated with laboratory analyses, decreasing subjectivity in visual diagnosis, and enabling more scalable, data-driven nutritional management.

Earlier studies tended to prioritize traditional artificial intelligence approaches, particularly classical ML methods, as observed in the work of Montes Condori et al. (2017) [13], mainly relying on manual feature extraction (color, texture, and shape), which represented the state of the art at that time. Subsequently, there was a shift toward conventional CNN models and the adoption of variants such as transfer learning, as reported by Ponce et al. (2021) [15], enabling robust feature extraction through the reuse of pre-trained models.

More recently, the incorporation of lightweight architectures and strategies focused on efficiency and deployment—such as compact models, edge inference, and mobile-device implementation—has gained prominence, alongside the integration of attention mechanisms and meta-learning to address severity levels and multi-task scenarios (deficiencies and diseases) within a single pipeline [14,16,49,86]. Despite this maturation, evidence suggests that the transition of these technologies remains incomplete, particularly between laboratory prototypes and solutions validated under real-world conditions [213].

Scientific production shows a geographic concentration in specific countries and institutions, notably India, reflecting the socioeconomic importance of agriculture and the demand for AI-based applications in regions with high production intensity. This is exemplified by authors such as Muthusamy and Ramu (2024, 2025a, 2025b) [14,117,199] from the Vellore Institute of Technology, who introduced advanced DL techniques. However, this asymmetry may introduce biases, as dominant crops, environmental conditions, and production systems in certain countries shape dataset characteristics, research priorities, and even experimental protocols.

Conversely, regions with high vulnerability and dependence on subsistence crops are underrepresented, limiting generalization and reducing the likelihood that solutions will emerge for these contexts. From both a social and scientific perspective, this gap represents not merely an “absence of articles,” but an indication that research remains misaligned with broader and more diverse global needs.

In Brazil, the University of São Paulo (USP) emerged as the most prominent institution, contributing three relevant articles to the analysis and consolidating its position as one of the most productive institutions in this domain. The application of AI for nitrogen diagnosis in maize was investigated by Romualdo et al. (2014) and Montes Condori et al. (2017) [12,13]. In addition, other authors followed this research line by applying crop-specific models and domain knowledge to nutritional diagnosis in maize and sorghum, contributing to advances in the field, as reported by Romualdo et al. (2014) and Martins et al. (2025) [12,23].

The predominance of studies focusing on certain crops, such as rice, and on macronutrients—particularly NPK—is justified by two main factors: (i) the economic and agronomic relevance of these crops and nutrients, and (ii) greater data availability, including more detailed technical documentation and visually detectable symptoms in RGB imagery. In contrast, agronomically important nutrients often associated with pest resistance, abiotic stress tolerance, and plant vigor—such as some micronutrients and beneficial elements—have been less explored due to their subtle or ambiguous visual symptoms.

Furthermore, symptoms overlap among nutrient deficiencies, phenological variability, and interactions with environmental stressors make visual diagnosis inherently ambiguous. This suggests that progress toward practical applications depends less on developing “more models” and more on generating stronger empirical evidence through realistic experimental designs, standardized data collection and validation protocols, and datasets that closely reflect real-world variability.

A major barrier to progress in this field is the limited sharing of data—and frequently code and experimental details—which reduces reproducibility, hinders model comparison, and fosters a landscape of “high-performance” results that often fail to generalize beyond specific datasets. The impact extends beyond transparency; without standardized and publicly available datasets, it is difficult to measure true model progress, establish benchmarks, or rigorously evaluate generalization.

Based on the observed data, greater data sharing occurred primarily in large datasets, suggesting that data collection capacity and institutional incentives influence openness. From a practical standpoint, future reviews and journals should encourage minimum requirements for data description, validation protocols, and data sharing whenever feasible.

Regarding models and architecture, there is a clear trend toward increased use of DL, consistent with its advantage in efficiently learning complex visual patterns and features. Nonetheless, classical ML models can achieve competitive performance in specific scenarios when properly optimized, particularly in contexts with limited data or well-defined features that effectively separate classes (Romualdo et al., 2014) [12]. Therefore, model selection should be driven by application requirements rather than average performance alone; computational cost, robustness to field variability, interpretability, and deployment feasibility are key considerations.

Within this context, pre-trained architectures such as ResNet, VGG, and MobileNet play a central role by balancing performance and knowledge reuse, while lightweight models stand out for deployment on mobile devices and embedded platforms, as demonstrated in several studies [16,101,214].

Although many studies report high performance, results must be interpreted with caution, as computer vision applications in agriculture are particularly sensitive to (i) small datasets; (ii) class imbalance; (iii) improper data partitioning or data leakage; and (iv) lack of field testing. Consequently, high accuracy values do not necessarily indicate robust operational performance and may reflect publication bias or favorable experimental scenarios.

Overall, the results reported in the literature (n ≈ 179) show generally high-performance metrics, with most studies reporting values above 90%, several exceeding 95%, and some achieving over 99% in at least one metric (e.g., Accuracy, Precision, Recall, F1-score). Few cases reported lower values, suggesting a degree of methodological maturity. However, performance comparisons across studies remain challenging, as metrics vary according to task type: classification studies report Accuracy, Precision, Recall, and F1-score; detection studies report mAP@IoU; and regression or concentration-estimation studies use R², RMSE, and MAE.

Accuracy alone is insufficient to assess model performance, particularly in imbalanced datasets where results may be inflated. Accordingly, F1-score and Recall are more informative for classification, mAP@IoU for detection tasks, and R² combined with RMSE/MAE for regression. Higher metric values are often reported in proprietary datasets, controlled environments, and simple hold-out splits. When public datasets and reproducible partitions are lacking, the risks of bias, data leakage, and overfitting increase, underscoring the importance of standardized reporting, transparency, and external validation.

To enable practical application under real-world conditions, several priorities emerge:

• Field validation using minimum data collection protocols, including environmental conditions, lighting, varieties, phenology, devices used, and external testing.
• Data and code standardization and sharing, with comprehensive documentation of the project pipeline.
• Lightweight models with interpretable outputs, particularly for edge/mobile deployment, incorporating strategies for multiple deficiencies and severity levels that reflect producers’ realities.

Although several reviews address computer vision applications in agriculture, many adopt a broad perspective encompassing stress diagnosis, disease detection, and crop monitoring, thereby diluting the specific challenges of nutritional deficiency diagnosis. In contrast, this systematic review focused on a comparative synthesis centered on foliar nutritional deficiencies, integrating three key dimensions commonly addressed separately in other reviews: (i) model families (ML, DL, or hybrid) and architectures; (ii) image modalities and acquisition platforms; and (iii) validation protocols and evidence of generalization, including controlled versus field environments, operational indicators, and reproducibility through data provenance and sharing. The integration of these dimensions highlights the relevance of evidence for real agricultural scenarios, where the current bottleneck is not merely achieving higher performance.

This review identified structural patterns that explain why practical adoption remains limited, including the focus on visually detectable crops and nutrients, the predominance of proprietary and non-public datasets, and limited standardization of data collection and validation protocols, all of which constrain cross-study comparison. Accordingly, this review contributes a critical assessment of result validity conditions, mapping both advances (architectures, mobile applications, efficiency) and persistent weaknesses (transparency, generalization, and field evaluation).

The practical implications of these findings support the transformation of prototypes into reliable tools by prioritizing field validation, protocol standardization, and data and code openness whenever possible, thereby enabling benchmarking and cross-study comparison.

Likewise, there is a clear opportunity to develop research that addresses producers’ needs more directly, including simultaneous assessment of multiple deficiencies, severity levels, and lightweight models suitable for mobile deployment. These elements connect the scientific synthesis of this review with a practical pathway toward real-world adoption.

In summary, although there has been rapid progress in architecture and reported performance, structural bottlenecks related to data availability, reproducibility, and validation persist. Consistent methodological practices and publicly available benchmarks are essential to bridge the gap between academic results and reliable, field-ready technologies in agronomic contexts.

5. Conclusions

This systematic review analyzed the literature from 2012 to 2026 on the use of computer vision combined with ML and DL for the identification and classification of foliar nutritional deficiencies, revealing recent growth and a predominance of DL-based models, particularly those using RGB images and transfer learning architectures. This study critically organized evidence demonstrating the integration and comparison of models, architectures, image modalities, acquisition platforms, and validation procedures, highlighting reported performance results to assess model robustness and to identify methodological weaknesses, especially for field applications.

According to the analyzed results, advances in architecture and performance have not been accompanied by equivalent progress in reproducibility practices and external validation, limiting cross-study comparison and the generalization of architectures to real-world scenarios. This indicates that the practical implementation of these technologies depends less on “new networks” and more on well-structured, diverse, and well-documented datasets, clear data-partitioning criteria, evaluation under real-world conditions, and methodological transparency.

Thus, progress toward field deployment must involve external testing and realistic environmental variability to generate consistent evidence and results, alongside standardization and data sharing, and an expanded agronomic scope beyond primary macronutrients such as NPK to include micronutrients, multiple deficiencies, and severity levels. From an application standpoint, lightweight and interpretable models are essential for deployment on mobile devices and edge-based solutions.

Despite the high performance reported for ML and DL approaches, robustness outside controlled scenarios remains limited. Practical consolidation therefore depends on external validation, methodological standardization, and data transparency to enable meaningful comparisons across studies.

In conclusion, ML and DL offer promising potential for more automated and efficient nutritional diagnosis; however, practical adoption requires raising the standard of evidence through improved data quality, more realistic validation, and methodological transparency. This shift is essential to transform the advances reported in the literature into reliable and consistent real-world outcomes.