Can Yield Prediction Be Fully Digitilized? A Systematic Review

Abstract

Going beyond previous work, this paper presents a systematic literature review that explores the deployment of satellites, drones, and ground-based sensors for yield prediction in agriculture. It covers multiple aspects of the topic, including crop types, key sensor platforms, data analysis techniques, and performance in estimating yield. To this end, datasets from Scopus and Web of Science were analyzed, resulting in the full review of 269 out of 1429 retrieved publications. Our study revealed that China (93 articles, >1800 citations) and the USA (58 articles, >1600 citations) are prominent contributors in this field; while satellites were the primary remote sensing platform (62%), followed by airborne (30%) and proximal sensors (27%). Additionally, statistical methods were used in 157 articles, and model-based approaches were utilized in 60 articles, while machine learning and deep learning were employed in 142 articles and 62 articles, respectively. When comparing methods, machine learning and deep learning methods exhibited high accuracy in crop yield prediction, while other techniques also demonstrated success, contingent on the specific crop platform and method employed. The findings of this study serve as a comprehensive roadmap for researchers and farmers, enabling them to make data-driven decisions and optimize agricultural practices, paving the way towards a fully digitized yield prediction.

1. Introduction

Estimating crop production is a crucial component in agriculture and has proven to be an effective approach for addressing food security concerns [1]. The World Health Organization [2] estimates that 820 million people worldwide still have insufficient access to food, while the Food and Agriculture Organization (FAO) projects a 70% increase in food demand required to support the global population of 9.1 billion by 2050 [3]. The droughts, floods, and heatwaves brought on by climate change are also putting added pressure on food production in many regions of the world [3]. In this context, yield prediction is an essential strategy that empowers farmers and the agricultural industry to manage resources efficiently, make informed decisions, and plan the harvesting, storage, processing, and logistics operations of the production, leading to increased productivity and cost savings. Moreover, timely forecasts enable farmers to plan for potential risks, such as severe weather events or pest outbreaks, allowing them to take prompt action and mitigate their impact [4]. Nevertheless, the estimation of crop production is a complex and intricate process that depends on a multitude of factors, such as the microclimate, weather, soil characteristics, fertilizer usage, and seed variety [5]. Therefore, numerous methods and techniques have been developed and used for optimizing yield prediction and improving the effectiveness of the developed models [6]. Precision agriculture could play a key role to yield estimation by utilizing various sensors including satellites, drones, and ground-based sensors, transforming the process of yield prediction by generating a plethora of data [7].

There have been a number of review papers focusing on the use of smart farming for yield prediction which offer insights into the challenges and opportunities of using remote sensing for crop management [8,9,10]. Several of these articles focus on the yield prediction of particular crops that are widely grown, such as maize, rice, sugarcane, sugar beet, and vines [11,12,13,14,15], while others include a more general overview of remote sensing technologies for specific application domains, such as crop management, crop monitoring, phenology, and other ecophysiological processes [16,17,18,19]. As reported [20], the relationship between vegetation indices obtained from remote sensing images (proximal, Unmanned Aerial Vehicles—UAVs, satellites) and crop yield is not static, but varies by vegetation stage. Towards this direction, a review of 69 studies by Benos et al. [21] highlighted a number of prediction levels at a specific vegetation stage or time before harvest. Schauberger et al. [22] performed a systematic review of crop yield forecasting methods in three often-used data domains: weather, remote sensing, and crop mask. By reviewing a large database (covering more than 350 articles), they reported that the most commonly-used models include statistical, process-based, and machine-learning models.

In relation to machine-learning models, the growing adoption of AI has allowed a noteworthy rise of studies focused on yield prediction [23,24,25]. Machine learning (ML) models treat the output, the crop yield, as an implicit function of the input variables, such as weather components and soil conditions, which can be very complex [26]. Many studies have used supervised and unsupervised learning, including various analytical models like Decision Trees, Random Forest, Support Vector Machines, Bayesian Networks, and Artificial Neural Networks [26,27,28]. Even though several review papers deliver a narrative overview of the topic [29,30,31], limited studies examine in depth all the necessary aspects for yield estimation. In this context, Van Klompenburg et al. [32] provided a systematic review of ML methods in yield prediction, including 567 relevant studies from six electronic databases. According to their findings, the algorithms that are most widely used were Neural Networks (NN) and Linear Regression algorithms, followed by Random Forest (RF) and Support Vector Machines (SVM). The most applied deep learning (DL) algorithm is Convolutional Neural Networks (CNN), and the other widely-used algorithms are Long-Short Term Memory (LSTM) and Deep Neural Networks (DNN). These findings are aligned with the systematic review of Oikonomidis et al. [33], who also reported the rapid increase of DL methods in crop yield prediction over the last five years. Similarly, the systematic review conducted by Muruganantham et al. [20] concluded that the performance and accuracy of the DL approach for crop yield prediction are better when compared to traditional ML approaches. Nevertheless, they are difficult to train and need recently developed hardware and optimization methodologies [34]. Large amounts of data are required to achieve good accuracy, and the complexity of DL approaches increases the algorithm’s time complexity [35]. When assessing ML techniques for achieving high levels of prediction performance, special attention should be given to different scales. Although prediction models at the regional scale could exhibit good accuracy, their usefulness to inform the decision-making of individual farmers might be severely limited according to the systematic review of Leukel et al. [36]. The review also accentuated the greater effort required for collecting field-level yield data (e.g., in-field sampling) compared with accessing readily available yield data from governmental bodies and regional associations. Wang et al. [33] also evaluated the applicability of DL for yield prediction on multiple scales and listed some representative studies regarding the nature of application and performance.

Although research has made great strides and crop yield prediction models can estimate the actual yield reasonably, better model performance is still desirable [37]. In the pursuit of enhancing agricultural productivity and ensuring food security, there is a pressing need for further advancements in yield prediction techniques. To address this requirement, this study aims to conduct a comprehensive systematic literature review, focusing on the deployment and integration of cutting-edge technologies such as satellites, airborne, and ground-based sensors in the context of crop yield prediction. By synthesizing this knowledge, we aim to provide valuable guidance for researchers, policymakers, and practitioners in the agriculture sector to make informed decisions and develop improved crop management strategies. Specifically, our review goes beyond previous work by combining multiple aspects of the topic including crop types, key sensors and platforms, data analysis techniques, and their respective performance for estimating yields. To this end, the following research questions are developed to guide the study:

• Which countries have been the key contributors to research related to the deployment of satellites, airborne, and ground-based sensors for crop yield prediction?
• Which crop types have been predominantly used for yield estimation in the context of remote sensing technologies?
• What are the most commonly employed remote sensing platforms and data analysis techniques for predicting crop yields in the existing literature?
• Among the various methods and platforms utilized, which ones have demonstrated better performance and accuracy in predicting crop yields?

By answering the above questions, this paper aims at providing a comprehensive and objective framework of the topic. It also identifies gaps in the existing research, and highlights hotspots where further investigation is needed in this rapidly growing field.

2. Materials and Methods

2.1. Scientific Article Search

In this study, peer-reviewed articles related to the application of remote sensing technologies in yield prediction were extracted, aiming to identify relevant studies from the earliest instances to the present day. To this end, a systematic search procedure was developed by utilizing Scopus “www.scopus.com (accessed on 1 February 2023)” and Web of Science (WoS) “www.webofscience.com (accessed on 1 February 2023)” search engines following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework [38]. Specifically, the PRISMA Statement consists of a 27-item checklist and a four-phase flow diagram, aiming at helping authors improve the reporting of systematic reviews and meta-analyses [38]. To ensure a comprehensive selection of relevant research articles for the analysis, the study’s approach was designed based on framed research questions and the aim of the review. It was acknowledged that using “yield prediction” alone as a search string would have generated a large number of published articles from various application fields that were not likely related to the aim of the review, leading to a complicated search. Therefore, the research words have been deliberately chosen, also considering relevant systematic reviews [22,32,39] to narrow down the focus from a main concept to a central idea. Specifically, the query used for encompassing all the works related to the topic without risking excluding any item is presented in Table 1.

Table 1. Search engines and queries that were used for the scope of this study.

Search Engine	Query
Scopus	TITLE-ABS-KEY (“yield forecasting” OR “yield prediction” OR “yield estimation” OR “crop modeling”) AND TITLE-ABS-KEY (“satellite” OR “UAV” OR “proximal” OR “remote sensing” OR “proximal sensing” OR “aerial”)
WoS	TS = (“yield forecasting” OR “yield prediction” OR “yield estimation” OR “crop modeling”) AND TS = (“satellite” OR “UAV” OR “proximal” OR “remote sensing” OR “proximal sensing” OR “aerial”)

Then, a filtering step was conducted by exploiting the exclusion criteria directly available in the Scopus and WoS search engine, that is, document type, language, and publication year. Open-access articles published in the English language were only selected, while review articles and conference papers were excluded. This was based on the fact that open-access publishing adheres to the principles of open science, fostering transparency and ensuring that research is readily accessible for thorough examination, and thereby upholding the fundamental tenets of scientific integrity. Furthermore, the time span of the investigation encompassed the entire body of literature from 2002 to 2022.

The search query generated 725 records through Scopus and 704 through WoS, with publication data containing information on the “Author, Title, Source”, “Abstract, Keyword, Addresses”, and “Cited, References and Use” categories, organized into fields. Moreover, by removing the repeated and review articles across the two selected databases, 864 articles were screened by title and abstract.

2.2. Article Selection Criteria

The initially retrieved articles were chosen based on specific criteria, including the type of remote sensing technology utilized in the study and the method employed for yield prediction. Analyzing the abstracts of these articles aided in identifying relevant keywords and assisting in the article selection process. To ensure the relevance and focus of the review, the following exclusion criteria were applied:

• Records not pertinent to the research objective (e.g., satellite RNA in plant pathology) were excluded;
• Articles falling within the agricultural sector but not directly related to crop yield prediction were also removed from consideration;
• Publications that did not incorporate the use of satellites, airborne, or ground-based sensors for crop yield prediction were excluded;
• Literature search for articles that are published between 1 January 2002 to 31 December 2022;
• Articles were included only if they forecasted crop yield, either in absolute or relative terms, and provided performance metrics for evaluation. In order to ensure consistency and comparability, particular attention was given to the presence of evaluation metrics such as R² (the coefficient of determination) and error metrics like the Root Mean Square Error (RMSE). Studies lacking these metrics were omitted from the dataset to standardize the evaluation process.

After applying all the exclusion criteria, a total of 456 full text articles were assessed for eligibility. Figure 1 presents the process for article selection and rejection from databases, based on the PRISMA framework.

Figure 1. Systematic review procedure for article selection.

The eligibility process involved thoroughly analyzing the full articles to ensure that only the studies that met the necessary aforementioned criteria were included. As a result, a total of 269 studies were deemed suitable and incorporated into this comprehensive review.

2.3. Scientific Studies Classification & Statistical Analysis

The selected papers were tabulated and standardized to enable comparison and systematic evaluation by extracting the following variables from each study:

• Study data: lead author, year, title, citations;
• Experiment setup: study region, type of crop;
• Platform type: Satellite, Airborne Measurements (Unmanned Aerial Systems—UAS or Manned Flight), Ground based Measurements;
• Method type: machine learning, statistical analysis, model-based approach, Vegetation Indices (VIs);
• Evaluation: performance measures (e.g., R², RMSE, MAE).

Subsequently, the actual data collected from the papers were subjected to statistical analysis using XLSTAT software version 2016 from Addinsoft (www.xlstat.com, accessed on 1 April 2023). This analysis involved determining the number of research articles produced annually and by type. Additionally, further analyses were conducted based on crop type, platform type, sensor type, and the method’s focus area for each year over the past two decades.

3. Results and Discussion

One of the principal findings of this study pertains to the number of publications per year from 2002 to 2022, which sheds light on the evolving trends and research activity in the field of yield prediction using remote sensing technologies. According to Figure 2, from 2002 to 2012, the publication rate was low, with an average of roughly one paper per year. However, between 2013 to 2017, the publication rate increased to an average of approximately six papers annually, indicating a growth stage. From 2018 onwards, a rapid increase in publications is evident, confirming the growing interest among researchers, which also reflects the yield prediction used in the literature. Specifically, the number of publications surged from 15 in 2018 to 42 in 2020 and reached its peak at 92 in 2022.

Figure 2. Number of publications per year throughout the period 2002 to 2022.

The higher number of articles in the last years can be explained by a confluence of factors, such as technological advancements in the Information and Communications Technology (ICT) area, augmented research funding, and an expanding understanding of remote sensing applications.

The list of selected papers for the review is summarized in Appendix A, Table A1, which includes relevant information such as the Title, Crop, Method, and Platform used in each study. This comprehensive summary allows readers to access and refer to the key details of the selected papers efficiently, aiding in the understanding and evaluation of the research conducted for the review.

3.1. Key Contributor Countries

This systematic review also provided insights into the geographical distribution of research and the key contributors in the field. Specifically, studies have been conducted in 55 countries (Figure 3), with China most frequently appearing, followed by the USA, India, Australia, and Brazil. There are also many experiments in developing countries, but often only with a single study on a single crop. Forecasting efforts in Europe are spread out geographically, largely following country size and production share, with a dearth of studies particularly in Eastern Europe. It is important to highlight that these findings are related to the study areas within the articles, not the countries of authorship.

Figure 3. Top 10 countries in terms of publications 2002–2022.

The most active country in terms of experiments for the whole period encompassed by this study is, clearly, China with over 93 publications. Following closely, the USA ranks second with 58 publications. India and Australia occupy the third and fourth positions, respectively, with 11 research studies each, while Brazil closely trails with 10 studies. In a more detached group, the majority then consists of European countries (Germany, Spain, Italy, France) with <8 publications.

The number of citations received is often used as a proxy for research quality. However, it should be noted that this metric alone may not provide a completely accurate representation, as various factors, including the research institute, the researchers’ country of origin, and the target audience, could influence citation counts [40]. Figure 4 illustrates the impact of research from different countries, and it becomes evident that China and the USA stand out, outperforming other countries in terms of citations. Notably, European countries, such as Germany and Spain, follow at a considerable distance with less than 370 citations, while Australia and Brazil are positioned further down the ranking. It is essential to highlight that these rankings are based on the currently available information and may be subject to updates as more recent citations become accessible, potentially influencing the relative positions of the countries in the future.

Figure 4. Geographical distribution of citations for all the selected articles (2002–2022).

By examining publication patterns and citation metrics, it was possible to identify the countries that have made significant contributions to the topic of interest, helping researchers understand the global landscape of research and identify potential collaboration opportunities. It is evident that the USA and China have emerged as the most influential countries in the field of crop yield estimation using remote sensing technologies. These two (2) nations have demonstrated a significant presence with a substantial number of research articles focused on crop yield estimation, remote sensing applications, and related subjects. Moreover, their prominent position in terms of citations underscores their consistent production of high-quality research, substantial contributions to advancements in the field, and a profound understanding of effectively harnessing remote sensing data for accurate yield prediction. The notable impact of their research could be explained by the fact that they have the biggest economies and invest heavily in research and development. Consequently, they employ a large number of researchers who produce research publications [41].

3.2. Crops Used for Yield Estimation

The choice of crops for yield estimation is a pivotal aspect of research in the field of remote sensing-based agriculture. Through a thorough analysis of the literature, the study identified the most frequently studied crops used in yield estimation through remote sensing techniques. In total, the research encompassed a diverse array of crops, amounting to 48 different types, which were further classified into nine categories based on the Food and Agriculture Organization (FAO)’s classification [42]. Figure 5 illustrates the number of studies that included crops from each category and the prominent crops that have been extensively researched in the field of remote sensing-based yield estimations. Several studies addressed multiple crops, which means the total number of crops illustrated is greater than the number of studies analyzed.

Figure 5. Categories of crops included in literature between 2002 and 2022.

Wheat (including durum wheat), maize, and rice emerge as highly studied crops, not only within the cereal category, but also overall. Additionally, oilseed crops, with soybeans leading the way, also receive significant attention in scientific publications. On the other hand, the fruits and nuts category along with vegetables and melons appeared to be the least researched category in terms of publications. It is noted that the category “Grass crops” comprises various crops, including Bachiaria pastures, Grassland, Miscanthus, perennial bioenergy grass, and ryegrass. Similarly, the category of “tomato” also includes research on processing tomato crops (Figure 6).

Figure 6. Number of studies per crop category and crop.

Overall, the prominent crops that were commonly utilized for yield prediction included cereals and oilseed crops. These crops were selected due to their nutritional value and, therefore, their economic significance, data availability, and relevance to global food security [43,44]. Another key factor influencing their widespread use could be the availability of extensive datasets, encompassing historical yield records, agronomic practices, and weather data. Such data availability facilitates researchers in conducting comprehensive yield prediction studies with greater ease. Moreover, these crops do not exhibit complex structure-like vineyards and orchards that may affect remote sensing results [45]. The frequent application of agricultural practices like irrigation and pruning that are conducted in other crops such as vineyards and orchards, could also affect the interpretation of the remote sensing results [46]. As a result, researchers may face additional technical challenges and data processing requirements for these crops. In contrast, cereals and oilseed crops generally experience less interference from such practices, leading to more reliable and consistent remote sensing outcomes.

3.3. Remote Sensing Platforms for Yield Forecasting Used in the Literature

The literature on remote sensing platforms for crop yield forecasting is vast and diverse. Different remote sensing platforms have different advantages and limitations in terms of spatial resolution, temporal resolution, spectral resolution, radiometric resolution, coverage area, revisit frequency, data availability, data cost, and data processing requirements. Therefore, selecting the most suitable remote sensing platform for a specific crop yield forecasting application depends on several factors, such as the type of crop, the scale of analysis, the purpose of forecasting, the available resources, and the user preferences.

The results indicate that various remote sensing platforms were widely utilized for crop yield estimation, with many studies employing multiple platforms simultaneously. Notably, the majority of the reviewed studies (62%) utilized satellite remotely-sensed data to generate yield forecasts throughout the growing season. However, for small-scale studies conducted on experimental plots, ground-based sensors (27%) or airborne sensors (30%) were more commonly employed (Figure 7). Nonetheless, even in cases where multiple platforms were used, satellites remained the primary choice for crop yield estimation. This diverse usage of remote sensing platforms underscores their versatility and the benefits they offer in gathering essential data for crop yield forecasting across different spatial scales and agricultural contexts.

Figure 7. Remote sensing platforms for yield forecasting used in the literature.

Satellites play a crucial role in crop yield prediction by utilizing a diverse range of sensors to measure electromagnetic radiation reflected or emitted from the Earth’s surface. Equipped with these sensors, satellites enable the spatial and multitemporal monitoring of soil and crop characteristics at different growth stages, providing valuable data for yield estimation. Figure 8 depicts the most common satellite systems used for yield prediction. Among the satellites commonly employed for this purpose, the Moderate Resolution Imaging Spectroradiometer (MODIS) emerges as the most frequently used, followed by Sentinel-2, Landsat, and Satellite pour l’Observation de la Terre (SPOT). Additionally, Synthetic Aperture Radar (SAR) sensors have also been utilized, with Sentinel-1 being the most prominent one.

Figure 8. (a) Satellite platforms for yield forecasting used in the literature; (b) ground-based platforms for yield forecasting used in the literature; (c) airborne platforms for yield forecasting used in the literature.

Yield predictions could also be derived based on data recorded from airborne platforms. According to the findings of this study, out of the total 269 studies reviewed, 84 of them utilized airborne data for crop yield prediction, including four manned flights. Among these, 45 studies utilized multispectral cameras, 30 studies deployed RGB cameras, and 15 studies utilized hyperspectral data. The least commonly used sensors were the thermal and synthetic aperture radar (SAR). It is obvious that several studies deployed more than one sensor, indicating the integration of multiple data sources to improve the accuracy and comprehensiveness of crop yield prediction models. The diverse usage of these sensors underscores the significance of integrating different data types to capture various aspects of crop growth and health for more informed yield forecasting.

In the case of ground-based sensors, the instruments were grouped based on their functionalities and applications. Specifically, the canopy sensors and analyzers category encompassed instruments for Chlorophyll Measurement (SPAD), Crop Health, and Nutrient Management (e.g., GreenSeeker, NTech Industries, Ukiah, CA, USA and CropCircle, Holland Scientific Inc., Lincoln, NA, USA), as well as Spectral Analysis and Canopy Analysis sensors (e.g., Spectroradiometer, spectrometers, Li-Cor 2000 Plant Canopy Analyzer, Li-Cor, Lincoln, NE, USA). Local meteorological stations were extensively deployed, appearing in 39 studies, making them the most commonly used ground-based sensors. Following closely, canopy sensors were frequently employed in the research. However, thermal sensors and LiDAR/Laser scanner data were the least deployed among the ground-based sensor categories.

Summarizing the results, researchers primarily utilized satellite platforms to acquire the necessary data for their studies. Satellites, compared to the rest of the platforms, can cover large areas and provide high temporal resolution, while being cost effective [47]. Moreover, satellites can be used in multisource data integration, such as the integration of optical and SAR remote sensing [48]. These advantages can explain why the majority of the studies incorporated satellite remote sensing approaches.

Respectively, UAS encompasses high spatial ground resolution and the ability to provide flexible and timely surveillance. However, UAS surveys require the storage and management of large amounts of data and preprocessing, while the datasets generated are limited to those collected by the user [49]. Consequently, deploying UASs on a commercial scale involves significant expenses, encompassing equipment, data processing, and software costs, which can be a substantial investment for small-scale farmers [50,51]. On the other hand, proximal sensors present distinct advantages in terms of precision and cost-effectiveness in agriculture. Since most of these sensors are active, they are not as restricted by weather conditions. Due to the close proximity in which the data are collected, there is less atmospheric interference, leading to more accurate data as well as high spatial resolution [52]. Nevertheless, they also have limitations pertaining to coverage, data interpretation, maintenance demands, and initial expenses. Therefore, the evaluation of the specific needs and available resources is essential when contemplating the adoption of remote sensor technology.

3.4. Data Analysis Techniques for Yield Forecasting Used in the Literature

Analyzing remote sensing products for yield prediction involves a range of methodologies that encompass ML, DL, statistical, and model-based approaches. These methods leverage the power of remote sensing data to estimate and predict crop yields accurately.

Based on the findings of this study (Figure 9), a statistical analysis is the most prevalent method employed for crop yield prediction in the reviewed studies. Following the statistical analysis, machine learning (ML) and deep learning (DL) methods are also widely used for yield estimation. In contrast, model-based approaches are observed to be utilized less frequently. Statistical analysis techniques often provide straightforward and interpretable relationships between variables, making them a popular choice for analyzing and understanding the impact of different factors on crop yields. Machine learning and deep learning methods, on the other hand, excel at capturing complex patterns and relationships in large and high-dimensional datasets, which is particularly advantageous when dealing with remote sensing data.

Figure 9. Overview of the methodological approach in the studies considered.

One of the significant discoveries of this study is the prominence of the Random Forest algorithm, which appeared in 89 studies, making it the most commonly used approach for crop yield prediction. This is aligned with the findings of another systematic review by van Klompenburg et al. [32], which reported that Random Forest is one of the most used models along with Linear Regression and the Gradient Boosting Tree. Following closely, Support Vector Machine (SVM) was featured in more than 52 studies, while Linear regressors were utilized in over 30 studies. Both XGBoost and Partial Least Square Regression (PLSR) are also frequently utilized, with more than 20 and 11 studies, respectively. It is worth noting that the Lasso Regression is another commonly used regularization technique (>14 studies), employing an L1 penalty to encourage sparsity in the model, resulting in the selection of relevant features. Similarly, the Ridge regression (>eight studies) is a variation of Linear Regression that incorporates a regularization term to prevent overfitting and enhance model performance when addressing multicollinearity. These methods have garnered significant attention in various research studies and applications, demonstrating their efficacy and versatility in yield prediction. Moving on to Neural Networks, Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN) take the lead as the top-ranked approaches, with 16 and 13 studies, respectively.

An intriguing observation from the study is that there were only three studies that employed ML and DL approaches between 2007 and 2010, whereas the vast majority of these studies were published from 2017 to 2022. This significant increase in the use of ML/DL techniques in recent years indicates a growing interest and recognition of the power and potential of these advanced methods for crop yield prediction using remote sensing data.

Model-based approaches, though less prevalent in this context, offer valuable insights and predictions by simulating the entire crop growth process and its intricate relationship with the environment from an ecological physiology perspective. These models integrate various factors such as crop characteristics, soil conditions, climate, and management practices to comprehensively simulate crucial physiological processes, including crop respiration, photosynthesis, phenology, biomass accumulation, crop distribution, and ultimately estimate crop yields. In this systematic review, several model-based approaches appeared for crop yield prediction using remote sensing data. It is essential to emphasize that model-based approaches typically necessitate a range of inputs, making remotely-sensed weather and biomass data particularly valuable for obtaining temporal and spatial information on a large scale.

Among model-based approaches, the Decision Support System for Agrotechnology Transfer (DSSAT) model [53] stood out with 13 featured studies, providing valuable insights into agricultural management practices and crop responses to environmental conditions. The Simple Algorithm For Yield model (SAFY) and WOrld FOod STudies (WOFOST) model [54,55,56] were each present in seven studies, offering simulations of crop growth under water-limited conditions and diverse environmental scenarios, respectively. AQUACROP [57,58,59], used in four studies, focused on crop water productivity, evaluating yield responses to water availability and irrigation management. The Agricultural Production Systems Simulator (APSIM) model [60,61,62] was investigated in three studies, encompassing various aspects of crop growth and management. Additionally, the PROSAIL (Prospect and Sail) model, deployed in seven studies, served as a radiative transfer model, enabling the assessment of crop health through light interactions in vegetation canopies. While it does not directly generate yield predictions, it was employed in conjunction with other models (APSIM, WOFROST) to extract Leaf Area Index (LAI) values, which were then used to estimate biomass.

It is important to note that different crop models operate based on distinct driving factors. For example, WOFOST focuses on carbon dioxide (CO₂), water, and temperature effects on yield, while AQUACROP emphasizes the impact of water stress on crop growth and yield, making it effective for simulating irrigation scenarios. APSIM, being a process-based model, considers a diverse range of soil processes, in addition to water balance and nutrient transformations [63]. Moreover, researchers have explored the benefits of coupled models, which combine two or more models with different principles and types. This approach aims to overcome the limitations of individual models, while capitalizing on their strengths, resulting in an improved simulation accuracy, modeling system stability, and reduced operational costs. These advances in model-based approaches contribute to a deeper understanding of crop–environment interactions and aid in making informed decisions for sustainable agricultural practices.

Each approach offered distinct advantages and addressed specific research objectives, enabling the extraction of meaningful information from remote sensing data for crop yield estimations. Specifically, the Statistical Analysis and Machine Learning methods are often used in crop yield estimation due to their ability to handle complex nonlinear relationships in high-dimensional datasets, as well as known parametric structures and unobserved cross-sectional heterogeneity [64]. Additionally, the performance of Deep Learning methods may be inadequate due to the fact that they heavily rely on the quality of the extracted features [65]. Finally, the low use of model-based methods on crop yield prediction could be explained by their high requirements for data and computational resources, and on their low flexibility compared to the other methods [66].

3.5. Spectral Vegetation Indices

Among the numerous vegetation indices developed, several have gained widespread adoption due to their effectiveness and versatility. As indicated by the results (Figure 10), the Normalized Difference Vegetation Index (NDVI) emerges as the most commonly used Vegetation Index. This can be explained by the high correlation this index presented, with key yield variables such as above ground biomass, crop height, and Leaf Area Index (LAI) [67,68]. The NDVI is also the most well-documented spectral vegetation index in the literature, resulting in reliable and accurate estimates of crop health and productivity, which are crucial for yield prediction [69]. Following closely is the Enhanced Vegetation Index (EVI), an improved vegetation index that addresses some of the limitations of the NDVI, particularly in areas with dense vegetation or atmospheric interference. Additionally, the LAI and Green Normalized Difference Vegetation Index (GNDVI) are widely employed in the studies. Each index offers unique advantages and applications, depending on specific research or monitoring objectives. Researchers, agronomists, and environmental scientists rely on these indices to analyze vegetation dynamics, assess crop health, monitor land cover changes, and make informed management decisions.

Figure 10. Most widely used vegetation indices (VIs) for crop yield prediction.

It is evident that various vegetation indices have gained popularity for their effectiveness, with the Normalized Difference Vegetation Index (NDVI) being the most widely used. The NDVI’s strong correlation with key yield factors like biomass, crop height, and LAI contributes to its prevalence. The Enhanced Vegetation Index (EVI) is also popular, addressing the NDVI’s limitations in dense vegetation or atmospheric conditions. The Leaf Area Index (LAI) and Green Normalized Difference Vegetation Index (GNDVI) are frequently employed too. These indices aid researchers, agronomists, and environmental scientists in analyzing vegetation, assessing crop health, monitoring land changes, and making informed decisions.

3.6. Accuracy Performance per Crop Category

Assessing accuracy performance per crop category is crucial for understanding the effectiveness of different methods and platforms in estimating yields for specific crops, aiding in informed decision-making and optimizing agricultural practices. Consequently, the highest performance measures (R²) obtained for each study were extracted and organized into tables based on crop categories.

When comparing different methods in the case of sugar (Table 2), beverage, and spice crops, ML techniques exhibit high performance, as shown in the table (Table 2). Specifically, the Random Forest method stands out with a noteworthy RMSE of 1.51 t/ha and an R² value of 0.94. It surpasses other methods including the Classification and Regression Tree, Support Vector Regression, and K-Nearest Neighbor [70]. This finding is in line with the results obtained by Canata et al. [71], where RF regression outperformed Multiple Linear Regression (MLR) in predicting sugarcane yields. Similarly, Martello et al. found that the RF regression yielded superior results in predicting coffee tree yields [72].

Table 2. Reported method, platform, and R², for sugar, beverage, and spice crop category.

Crop	References	Method	Platform	R2
Sugarcane	[73]	Statistical	Satellite × Proximal	0.53
	[74,75,76,77]	Statistical	Satellite	0.55 to 0.8
	[70,78]	ML, Statistical	Satellite	0.87 to 0.94
	[71]	ML	Satellite	0.70
	[79]	Model based	Satellite	0.86
Coriander	[80]	Statistical	Satellite	0.81 to 0.87
Tea	[81]	ML	Satellite	0.68 to 0.71
Coffee Tree	[82]	Statistical, Model based	Satellite	0.64 to 0.69
	[72]	ML, Statistical	Satellite	0.88 to 0.93

Furthermore, the most common platform used was satellite systems, indicating encouraging prediction accuracies (R² = 0.87) and RMSE = 11.33 (t·ha⁻¹) when compared to actual harvested yields [78]. Additionally, the utilization of SAR-based yield prediction models have also proved the potential to assist and support sugar mill technicians in refining yield estimates [75]. Nevertheless, a study by Duveiller et al. [77] highlights that the estimation of sugarcane yield is influenced by various aspects, namely: (1) the way time is regarded (thermal or calendar); (2) the purity of the signal; (3) how the information is extracted from the time series (i.e., the type of metrics); and (4) the timing of when the information is available. These factors can explain the different range of R² values retrieved from satellites for yield prediction (Table 2). Moral et al. [76] suggest that the empirical NDVI model is the most suitable approach for estimating sugarcane yield at the field level due to its simplicity and high accuracy throughout the entire crop cycle. In contrast to linear, logarithmic, power, and exponential models, a separate study [74] demonstrates that the polynomial model exhibits a significantly improved performance.

In the context of model-based yield prediction, the findings indicate a medium to high performance, with R² values ranging from 0.64 to 0.86. This can be explained by the selection of the model. A study conducted in the USA compared three statistical models that incorporated remote sensing and weather data. Among these models, the SiPAR model demonstrated a superior yield prediction compared to the cumulated DNVI (CNDVI) and Kumar and Monteith (K–M) models [79].

In the crop category of Vegetables and Melons (Table 3), ML techniques have demonstrated a high performance as well, achieving an R² value of 0.90. Apart from the deployed method, the selection of the VI plays a crucial role in achieving an optimal performance. According to the study conducted by Suarez et al. [83], the optimal results were produced when the Renormalized Vegetation Index (RDVI), Soil Adjusted Vegetation Index (SAVI), and Optimized Soil Adjusted Vegetation Index (OSAVI) were the predictor variables (R² = 0.77), with the lowest σ (10.75 t/ha) achieved with RDVI. EVI2 also performed better (R²  =  0.55) than GNDVI (R²  =  0.29). Another study [84] focusing on processing tomato crops identified plant height and VIs during the early to mid-fruit formation period as significant variables for predicting shoot masses. Notably, the NDVI and Weighted Difference Vegetation Index (WDVI) were found to be significantly important for predicting tomato weight, while VIs one (1) month prior to harvest were significant in predicting fruit quantity.

Table 3. Reported method, platform, and R² for the Vegetables and Melons crop category.

Crop	Reference	Method	Platform	R2
Chinese Cabbage White Radish	[85]	Statistical	Airborne	0.66 to 0.90
Carrot	[83]	Statistical	Satellite	0.29 to 0.78
African Eggplant	[86]	Statistical	Airborne × Proximal	0.54 to 0.87
Table Beet	[87]	Statistical	Airborne	0.89
Tomato	[88] *	Statistical	Satellite	0.69 to 0.81
	[84,89,90] *	ML, Statistical	Airborne	0.70 to 0.90

* Processing Tomato.

Moreover, recent findings [88] suggest a strong correlation between the development stages of the primary canopy in processing tomatoes and their final yield. This correlation may indicate a crucial stage during which the crops undergo discernible changes that can be detected using satellite-derived data. Additional research demonstrates the possibility of predicting average tomato biomass and yield up to 8 weeks before harvest, as well as at the individual plant level up to 4 weeks prior to harvest [89]. By employing time-series phenotypic features derived from UAVs, researchers observed a strong individual correlation between these features and the actual yield. Linear Regression models produced high values (R² > 0.7) in this regard [90].

In the context of oilseed crops (Table 4), the utilization of satellite NDVI series, captured fifty (50) days prior to harvest, has proven to be a reasonably accurate approach for estimating sunflower yields [91]. Furthermore, the effectiveness of Evolutionary Product-Unit Neural Network (EPUNN) models has been explored, revealing a superior accuracy compared to linear SMLR models, both in the training set and generalization set [92]. In the case of rapeseed yield estimation, plot-level VIs and leaf-related abundance showed a strong correlation, with an R² value above 0.75. Among the tested VIs, multiplying the NDVI, Chlorophyll Index Red Edge (CIred edge), Transformed Vegetation Index (TVI), and SAVI by short-stalk-leaf abundance yielded the most accurate results for yield estimation in rapeseed [93]. When it comes to model-based methods [63], the WOFOST model in comparison with the coupled CASA-WOFOST model demonstrated a faster running speed in yield simulations while maintaining a similar accuracy. This makes the proposed CASA-WOFOST model suitable for large-scale assessments using high-spatial-resolution images to obtain accurate yield simulations. An investigation was conducted to assess the potential of multisensor optical and multiorbital SAR data for monitoring winter rapeseed crops using the SAFY agrometeorological model. The results demonstrated that the assimilation of both SAR-derived dry matter (DM) and the optically derived Green Area Index (GAI) allowed for better control of the model compared to using SAR or optical data alone. This integration notably improved the optimization of parameters governing dry matter partitioning into leaves and effective light-use efficiency [94]. Another crucial aspect in satellite-based crop yield estimation is the spatial and temporal resolution of the deployed satellites. As highlighted by Chen et al. [95], sparse time series of satellite remote sensing, caused by low-temporal-frequency and/or cloud contamination, pose significant challenges for accurate crop yield estimation at regional to national scales. To address this limitation, the blending of high-spatial-resolution but low-temporal-frequency images with low-spatial-resolution but high-temporal-frequency images was proposed. This approach aims to increase the temporal resolution, while preserving essential spatial details, potentially enhancing the accuracy of crop yield estimations.

Table 4. Reported methods, platforms, and R² for the Oilseed Crop category.

Crop	References	Method	Platform	R2
Groundnut	[96]	ML, Statistical	Satellite × Proximal	0.96
	[97]	ML/DL, Model based	Satellite × Proximal	0.68
Sunflower	[91]	ML	Satellite	0.90
	[80]	Statistical	Satellite	0.56
	[92]	ML/DL, Statistical	Airborne	0.43
	[98]	Statistical	Satellite	0.91
Olive Tree	[99]	Statistical	Airborne	0.97
Palm Oil	[100]	ML/DL	Satellite	0.82
Canola	[101]	Statistical	Airborne	0.82
	[95]	Statistical	Satellite	0.86
Rapeseed	[93]	Statistical	Airborne × Proximal	0.81
	[63]	Model based, Statistical	Satellite × Proximal	0.86
	[94]	Model based	Satellite × Proximal	0.82
	[98]	Statistical	Satellite	0.97
Soybean	[102,103,104]	ML/DL, Statistical	Satellite	0.87 to 0.90
	[98,105,106,107,108,109,110,111]	Statistical	Satellite	0.49 to 0.98
	[112,113]	ML/DL	Satellite	0.85
	[114]	ML	Satellite	0.61
	[115,116]	ML, Statistical	Satellite	0.86 to 0.90
	[117,118]	ML/DL	Airborne	0.72 to 0.66
	[119]	ML	Airborne	0.89
	[120]	Statistical	Airborne	0.74
	[121]	ML/DL	Satellite × Proximal	0.85
	[122]	ML, Statistical	Satellite × Proximal	0.82
	[123]	ML	Airborne × Proximal	0.97
	[124]	ML/DL, Statistical	Satellite × Proximal	0.67

It is not surprising to find numerous studies that involve soybeans in their research, as soybean is a widely cultivated and economically important crop. A study [111] comparing various spatial resolutions found compelling evidence in favor of higher resolution imagery over lower resolution options. The authors suggest selecting an NDVI resolution that matches or exceeds the current cropland mask resolution, taking into consideration factors such as computation cost. Notably, an interesting finding from another research study [122] is that county-scale models perform relatively poorly in field-scale validation (R² = 0.32), particularly in high-yielding fields. However, these county-scale models show a similar performance to field-scale models when evaluated at the county level (R² = 0.82).

In the Fruits and Nuts category (Table 5), orchard yield estimation has predominantly been conducted using proximal sensing and airborne sensing, or a combination of both along with satellite data. High-resolution satellite images have also been employed as a standalone method, achieving a satisfactory performance with an R²-value of 0.87 [125,126]. The high efficiency of these methods could be attributed to their reliance on visual counting and the utilization of high-resolution data, which enable accurate and efficient orchard production estimations. Few studies used above-ground remote sensing to estimate tree production. The correlation between tree production and remotely-assessed features is not generic, and has to be calibrated for each orchard and each year to include climate and site effects [127].

Table 5. Reported methods, platforms, and R² for the Fruits and Nuts crop category.

Crop	References	Method	Platform	R2
Vineyards	[125,128]	Statistical	Satellite × Proximal	0.42–0.87
	[129]	ML	Satellite × Proximal	0.79
	[130]	ML/DL	Proximal	0.91
	[131]	ML, Statistical	Proximal	0.86
Almond	[132]	Statistical	Airborne	0.84
	[133]	ML/DL, Statistical	Satellite × Airborne	0.71
Apple	[134]	ML/DL	Airborne	0.88
Jujube	[135,136]	Model based	Satellite	0.62 to 0.78
Mango	[126]	ML/DL, Statistical	Satellite	0.77
	[127]	ML, Statistical	Airborne	0.77

In relation to root tuber and other crops (Table 6), ML approaches are quite common, achieving a higher (>0.90) performance in terms of accuracy when compared to other methods. In cotton cultivation, developing efficient tools for precise yield estimation before harvest is crucial, and the UAV multispectral remote sensing system holds significant potential for rapidly, accurately, and economically assessing agricultural crop characteristics and yields. The connection between crop growth indicators like LAI and chlorophyll with canopy spectral reflectance allows spectral indices collected during the growing season to be utilized for crop yield estimation, given the correlation between yield and the amount of photosynthetic tissue. This enables wide-scale application, contrasting with traditional measurements of agronomic parameters such as LAI and chlorophyll [137]. Additionally, the feasibility of estimating cotton yield using low-altitude UAV imaging was verified in this study [138].

Table 6. Reported methods, platforms, and R² for the Root tuber and other crops category.

Crop	References	Method	Platform	R2
Potato	[139]	Statistical	Satellite	0.65
	[140]	ML, Statistical	Satellite	0.89
	[141]	ML	Satellite × Proximal	0.86
	[67]	ML	Airborne	0.83
	[142]	ML, Statistical	Proximal	0.72
	[63]	Model based, Statistical	Satellite × Proximal	0.86
Cotton	[143,144]	Statistical	Airborne	0.52 to 0.94
	[145]	ML/DL	Airborne	0.85
	[146]	ML/DL, Statistical	Satellite	0.67
	[147]	Model based	Satellite × Proximal	0.96
	[137]	Statistical	Airborne × Proximal	0.84
	[148]	ML	Airborne × Proximal	0.93
	[138]	ML/DL, Statistical	Airborne	0.97
	[149,150]	ML, Statistical	Airborne	0.77 to 0.91
Sweet Potato	[105]	Statistical	Satellite	0.68
Cassava Tuber	[151]	Statistical	Airborne	0.87
Ramie	[152]	Statistical	Airborne	0.66
Milk Thistle	[63]	Model based, Statistical	Satellite ×Proximal	0.86
Grassland *	[153]	ML	Airborne	0.87
	[154]	Statistical	Airborne	0.75
Perennial Ryegrass *	[155]	ML	Airborne	0.93
Perennial Bioenergy Grass *	[156]	Statistical	Satellite	0.88
Brachiaria Pastures *	[157]	ML	Satellite × Airborne	0.75
Miscanthus *	[158]	ML, Statistical, Model based	Airborne	0.79

* Grasses and other fodder crops.

Researchers used mixed data sources, including airborne satellites and proximal sensors, to gather information and insights about leguminous crops (Table 7). Some studies may solely focus on using ML or DL algorithms, while others might combine both approaches or incorporate statistical methods for enhanced accuracy and interpretability. In the study conducted by Minch et al. [159], efficient flight parameters were investigated to create successful models for determining canopy heights, specifically for alfalfa yield estimation. The researchers strongly recommend using a flight parameter within the range of 50–75°, as it is likely to yield optimal data for accurate canopy height estimation in alfalfa fields.

Table 7. Reported methods, platforms, and R² for the Leguminous crop category.

Crop	References	Method	Platform	R2
Alfa Alfa	[160,161]	Statistical	Satellite	0.72 to 0.94
	[162]	ML/DL	Airborne	0.87
	[159]	ML	Airborne	0.84
	[163]	Statistical	Airborne	0.64
	[164]	ML, Statistical	Satellite	0.93
Red Clover	[165]	ML/DL	Airborne	0.90
Chickpea	[166]	ML	Satellite × Proximal	0.92
Snap Bean *	[167]	ML/DL	Airborne	0.98
Peas	[160]	Statistical	Satellite	0.95
Beans *	[168]	Statistical	Airborne × Proximal	0.70
	[169]	ML	Satellite	0.54
	[170]	Statistical	Satellite × Proximal	0.84
Faba Bean	[171]	ML, Statistical	Airborne	0.72

* Included in beans.

The category of cereals encompasses a wide range of methods and platforms, prompting its separation into two tables: cereals (Table 8), and maize and wheat (Table 9).

Table 8. Reported methods, platforms, and R² for the cereal crop category.

Crop	Reference	Method	Platform	R2
Cereal	[172,173]	Statistical	Satellite	0.71
Barley	[95,160,174]	Statistical	Satellite	0.86 to 0.93
	[175]	Statistical	Satellite × Airborne × Proximal	0.70
	[176,177]	Model based, Statistical	Satellite	0.6 to 0.77
	[178]	ML/DL	Airborne × Proximal	0.929
	[179]	ML × Statistical	Satellite × Proximal	0.88
	[180]	ML, Statistical, Model based	Satellite	0.47
Oats	[175]	Statistical	Satellite × Airborne × Proximal	0.79
	[178]	ML/DL	Airborne × Proximal	0.929
	[181]	Statistical	Proximal	0.90
Millet	[105]	Statistical	Satellite	0.68
	[169]	ML	Satellite	0.40
Sorghum	[105,161,182]	Statistical	Satellite	0.25 to 0.81
	[183]	ML/DL	Satellite × Proximal	0.35
	[169]	ML	Satellite	0.44
Rice	[105,184,185,186,187,188]	Statistical	Satellite	0.56 to 0.97
	[114,189,190,191]	ML	Satellite	0.43 to 0.95
	[192,193]	Model based	Satellite	0.89 to 0.96
	[194]	ML, Model based	Airborne × Proximal	0.75
	[195,196]	ML/DL, Statistical	Airborne × Proximal	0.22 0.51
	[197,198]	ML, Statistical	Airborne	0.76 to 0.8
	[97,199]	ML/DL, Model based	Satellite × Proximal	0.75 to 0.86
	[200]	Statistical, Model based	Satellite	0.80
	[201]	ML/DL	Satellite	0.81
	[202]	ML/DL	Airborne	0.84
	[203]	Statistical	Airborne × Proximal	0.64
	[204]	ML, Statistical	Airborne × Proximal	0.83
	[205]	ML, Statistical	Proximal	0.86
	[206]	Statistical, Model based	Airborne	0.94
	[207,208]	Statistical	Satellite × Proximal	0.66 to 0.90
	[209,210,211,212]	Statistical	Airborne	0.74 to 0.83

Table 9. Reported methods, platforms, and R² for wheat and maize.

Crop	References	Method	Platform	R2
Maize	[213]	Statistical	Satellite × Proximal	0.87
	[214]	Statistical	Airborne × Proximal	0.83
	[215]	Statistical	Airborne	0.74
	[105,107,108,111,160,161,169,216,217,218,219,220,221,222,223]	Statistical	Satellite	0.46 to 0.99
	[224,225]	Model based, ML/DL	Satellite	0.85
	[226,227,228,229]	Model based	Satellite	0.68 to 0.83
	[230]	Model based	Airborne × Proximal	0.855
	[231]	Model based	Proximal	0.68
	[91,114,232,233,234,235,236]	ML	Satellite	0.43 to 0.92
	[237]	ML, Statistical, Model based	Satellite	0.59
	[115,238,239]	ML, Statistical	Satellite	0.48 to 0.91
	[121,240]	ML/DL	Satellite × Proximal	0.75 to 0.85
	[1,102,103,104,124,241,242]	ML/DL, Statistical	Satellite	0.70 to 0.92
	[243,244,245,246]	ML/DL	Airborne	0.57 to 0.93
	[247,248]	ML, Statistical	Satellite × Proximal	0.35 to 0.98
	[249]	ML	Proximal	0.7
	[250]	Model based, ML	Satellite × Proximal	0.58
	[251]	Statistical, Model based	Airborne	0.81
	[97]	ML/DL, Model based	Satellite × Proximal	0.75
	[252]	Statistical, Model based	Satellite	0.73
	[253]	ML, Model based	Satellite	0.76
	[254]	ML x Statistical	Airborne	0.80
Wheat	[80,95,107,111,160,161,174,219,255,256,257,258,259,260,261,262,263]	Statistical	Satellite	0.37 to 0.99
	[264]	ML/DL, Model based	Satellite	0.83
	[265]	ML/DL	Satellite	0.75
	[1,266]	ML/DL, Statistical	Satellite	0.72 to 0.78
	[176,267,268,269,270]	Model based, Statistical	Satellite	0.48 to 0.86
	[180,271,272]	ML, Model based	Satellite	0.55 to 0.75
	[115,273,274]	ML, Statistical	Satellite	0.72 to 0.89
	[25,114,234,275,276,277]	ML	Satellite	0.51 to 0.99
	[177,278,279,280,281,282,283,284,285,286,287]	Model based	Satellite	0.49 to 0.86
	[288,289,290,291,292,293]	ML/DL	Satellite	0.79 to 0.93
	[294,295]	Model based, Statistical	Proximal	0.698 to 0.77
	[296,297]	Statistical	Proximal	0.46 to 0.48
	[298,299]	ML/DL	Proximal	0.83 to 0.891
	[300]	Model based	Proximal	0.84
	[301]	ML, Statistical	Airborne	0.81
	[302,303,304]	ML/DL	Airborne	0.62 to 0.85
	[305]	Statistical	Airborne	0.70
	[306,307,308,309,310]	ML	Airborne	0.62 to 0.93
	[311,312,313]	ML/DL, Statistical	Airborne	0.59 to 0.84
	[314,315,316]	ML/DL, Statistical	Airborne × Proximal	0.83 to 0.93
	[178,317,318]	Statistical	Airborne × Proximal	0.73 to 0.929
	[319]	ML, Statistical	Airborne × Proximal	0.78
	[179,320]	ML, Statistical	Satellite × Proximal	0.83 to 0.88
	[321,322]	ML/DL, Statistical Model based	Satellite × Proximal	0.68 to 0.91
	[323]	ML/DL, Statistical	Satellite × Proximal	0.50
	[63,324,325,326,327]	Statistical, Model based	Satellite × Proximal	0.61 to 0.93
	[328]	ML	Satellite × Proximal	0.89
	[329,330]	ML/DL	Satellite × Proximal	0.63 to 0.86
	[331]	ML/DL, Statistical, Model based	Satellite × Proximal	0.77
	[332]	Model based	Satellite × Proximal	0.49
	[333,334]	Statistical	Satellite × Proximal	0.55 to 0.76
	[175]	Statistical	Satellite × Airborne × Proximal	0.79

In the table focusing on wheat and maize (Table 9), it becomes evident that these crops have received special attention in the literature. The number of research papers dedicated to studying wheat and maize yield prediction is higher compared to other cereals, indicating their prominence in agricultural research. Moreover, the utilization of diverse approaches in predicting the yields of wheat and maize is also noteworthy. Researchers have explored a wide range of methods and platforms, including various machine learning algorithms, statistical models, and remote sensing technologies such as UAV multispectral imaging and satellite data.

Upon close examination of the provided table (Table 9), it becomes evident that a definitive and uniform trend in the methodologies employed for yield prediction is lacking. However, maize and, secondarily, wheat, rice, and soybean have emerged as extensively studied crops through the application of machine learning techniques. This observation is in accordance with the insights documented by Benos et al. [21]. The authors also reported that UAVs are constantly gaining ground against satellites mainly because of their flexibility and ability to provide images with high resolution under any weather conditions. Satellites, on the other hand, could supply time-series over large areas. At the same time, the range of approaches utilized aligns with a prior study [22] that has also observed a variety of methods used in predicting yields for staple crops, emphasizing that each new setting requires appropriate validation.

This information can be valuable for policymakers, farmers, and researchers to make informed decisions, optimize agricultural practices, and address food security challenges in an ever-changing climate and agricultural landscape.

Overall, the results emphasize the importance of assessing an accuracy performance for specific crop categories to enhance yield estimation methods. The highest performance measures (R²) from various studies were compiled into tables based on crop categories. Machine Learning (ML) techniques, particularly Random Forest, excel in predicting sugar, beverage, and spice crops. Satellite systems, including the Synthetic Aperture Radar (SAR), prove effective for sugarcane yield prediction. In vegetables, ML methods give promising results, considering key vegetation indices. Orchards benefit from proximal and airborne sensing, while leguminous crops are studied using a mix of ML, DL, and statistical methods. Wheat and maize receive extensive attention, employing diverse methods including ML, DL, statistical, and model-based approaches.

Finally, it is necessary to pinpoint an important constraint of this study. Due to the extensive volume of articles analyzed and the diverse methodologies employed therein, specific details concerning the performance assessment were not documented. These details include whether cross-validation was utilized, whether the testing dataset was segregated, or if the same dataset was employed for both training and model validation.

4. Conclusions

By employing this systematic approach to data analysis, the study aims to provide valuable insights into the trends, patterns, and contributions of different methodologies and technologies in the field of crop yield prediction using remote sensing tools.

Understanding the geographical distribution of research efforts and the significant academic institution in this domain is crucial for comprehending the research landscape. Our research revealed that China (93 articles with over 1800 citations) and the USA (58 articles with over 1600 citations) are key contributors to the field of crop yield prediction using remote sensing techniques. Based on the results, cereal crops (185 papers) emerged as the most commonly studied for yield estimation with wheat being the most predominant crop. Among the remote sensing platforms, satellites (62%) were the most frequently employed platforms followed by airborne (30%) and proximal sensors (27%). The study extensively evaluated various algorithms and models for predicting crop yields based on remote sensing data. In terms of methodologies, machine learning was featured in 142 articles, while deep learning was employed in 62 articles for the purpose of yield prediction. Furthermore, statistical methods were utilized in 157 articles, and model-based approaches were featured in 60 articles as mechanisms for predicting crop yields. The performance of machine learning and deep learning methods has shown high accuracy in crop yield prediction, while other techniques have also demonstrated success depending on the crop and method. These insights offer a comprehensive understanding of the research domain and could guide future advancements in remote sensing-based crop yield estimations.

By consolidating and analyzing data from multiple studies, our research contributes to a comprehensive understanding of the current state of remote sensing-based crop yield estimation. This synthesis helps identify trends, gaps, and areas of progress in the research domain, providing valuable guidance for future studies in this area. The identified influential countries, methodologies, and successful algorithms can serve as a foundation for designing more effective and targeted research.

The findings of this study hold the potential to advance the accuracy and applicability of remote sensing-based crop yield estimation techniques. This, in turn, could contribute to improved agricultural management practices, increased food security, and sustainable agriculture. The comprehensive overview provided by this research empowers the scientific community to make informed decisions and develop innovative approaches to further enhance the accuracy and utility of remote sensing-based crop yield estimation methods in the future.