Predictive Models Based on Artificial Intelligence to Estimate Crop Yield: A Literature Review

Abstract

This article presents a detailed review of methodologies for estimating crop yields in the context of growing global concern for food security and agricultural sustainability, with the main objective of analyzing, synthesizing, and comparing recent studies that apply artificial intelligence to yield prediction, identifying their strengths, limitations, and emerging trends. Approaches that integrate climatic variables, soil conditions, and agricultural management practices are examined. Artificial intelligence techniques, such as machine learning and neural networks, are effective at building robust predictive models. In several reviewed studies, these methods have achieved coefficients of determination (R²) greater than 0.85 and error reductions of 15% to 30% compared to traditional statistical approaches, confirming their high predictive potential. These models consider key elements such as temperature, precipitation, soil fertility, and agronomic decisions related to planting, crop choice, and fertilizer use. The article also discusses the challenges associated with model calibration and selection, given the complexity of agricultural systems and the variability of available data. The review covers studies published between 2016 and 2024, a period in which there has been a notable advance in the application of hybrid and deep learning approaches in the agricultural field. The importance of further research into hybrid approaches that integrate various techniques to improve prediction accuracy is highlighted. Finally, the strategic role of artificial intelligence in agricultural decision-making, in promoting sustainable practices, and in strengthening global food security is underlined.

1. Introduction

The food supply is a fundamental pillar for global well-being, but the agricultural sector faces multiple constraints that are clearly reflected in society. Malnutrition remains a persistent problem in various regions of the world. According to data from the Food and Agriculture Organization of the United Nations (FAO), approximately 811 million people worldwide suffered from chronic malnutrition between 2020 and 2022 [1]. By 2050, the world’s population is expected to reach 9.7 billion people, requiring a 50% increase in food production from 2012 levels. This challenge is compounded by factors such as climate change, soil degradation, and the urgency of modernizing agricultural practices to achieve sustainable production and ensure food security. Climate variability and extreme weather events represent major threats to global agricultural productivity, reducing the stability of food systems and challenging sustainability goals. According to recent FAO reports (FAO, 2021; FAO, 2023), approximately 22% of global yield losses are directly attributed to adverse weather conditions such as droughts, floods, and temperature extremes. These factors underscore the need for predictive models capable of anticipating yield fluctuations under changing climatic scenarios [2,3].

The lack of accurate, timely information on predicted agricultural yields limits the efficiency of on-farm decision-making, leading to economic losses and challenges to food security. By incorporating data on climate, soil characteristics, and agricultural practices, yield prediction can strengthen planning and decision-making in the agricultural environment [4,5]. Faced with the challenges posed by population growth, climate change, and the scarcity of natural resources, it is essential to have effective predictive tools that enable anticipating and mitigating impacts on food production [5,6].

Climatic variables, such as temperature, precipitation, and solar radiation, play a crucial role in determining crop growth and productivity. Similarly, soil quality considering its structure, pH, and nutrient levels directly affects yields. Agricultural practices, such as fertilizer application, efficient irrigation, and planting methods, also significantly impact the results achieved [7,8,9]. In this scenario, artificial intelligence (AI) has emerged as a promising alternative for predicting agricultural yields. This technology can identify complex patterns and correlations within extensive climate, soil, and agricultural management datasets, facilitating more accurate projections of production results [10,11]. Having advanced performance estimates offers multiple benefits. It enables producers to strategically plan their planting and harvesting activities, optimize resources, and select crops effectively, thereby reducing costs and increasing profitability. In addition, these predictions are critical for global food security, as they enable the identification of production deficits and the implementation of preventive measures to ensure an adequate food supply [12].

This review focuses on the application of artificial intelligence to predict crop yields, considering climatic variables, soil nutrition, and agricultural practices. The main datasets, the variables used, and the AI strategies applied in this field are presented.

2. Materials and Methods

The methodological protocol adopted for this literature review was based on the guidelines established by [13], the phases of which are illustrated in Figure 1. These guidelines structured the process of collecting, selecting, and analyzing academic information. For the initial search, a strategic combination of keywords related to agricultural yield prediction was employed, including terms such as “crop yield”, “machine learning”, “artificial intelligence”, “remote sensing”, “soil fertility”, and “climatic factors”; Boolean operators (AND, OR) were sed to refine the query and capture the most relevant studies. The bibliographic exploration was conducted in recognized scientific databases, including Scopus, Web of Science, and ScienceDirect [13]. This process followed a systematic approach aimed at ensuring transparency, replicability, and comprehensiveness in data collection; inclusion and exclusion criteria were established to select only empirical studies and reviews published between 2016 and 2024, with an emphasis on those applying artificial intelligence techniques to the analysis of agricultural performance. Furthermore, a data filtering and synthesis procedure was employed to identify methodological patterns, emerging trends, and gaps in the literature. Figure 1 presents the general outline of the methodological process followed for this review.

To include as many relevant investigations as possible, an advanced search strategy combining Boolean operators and truncation terms was created; using this strategy, specific inclusion criteria were set to select the studies for analysis. Only articles presenting original research, published between 2016–2024, and addressing crop yield prediction with artificial intelligence techniques linked to climate variables, soil, and agricultural practices were included: bibliographic reviews were excluded (some are discussed as related work in the results section). For each selected article, key metadata such as title, authorship, publication year, journal, and main findings on agricultural yield prediction were extracted. During the planning stage, several guiding research questions were formulated to direct the analytical process and assess how artificial intelligence has been applied to crop yield prediction: (1) What are the most influential climatic factors affecting agricultural yield, and how are they integrated into AI-based predictive models? (2) Which soil quality features are most incorporated into these models? (3) Which agricultural management practices most significantly influence crop productivity? (4) What are the predominant artificial intelligence approaches used to estimate agricultural yields?

The search strategy was restricted to key concepts relevant to the study’s focus. Given the multiple interdisciplinary applications of artificial intelligence, it was anticipated that some articles would not be relevant to the specific purpose of this review; therefore, the abstracts of the retrieved articles were reviewed to identify synonyms and related terms to the initial keywords. The search was conducted across five academic databases. To minimize the risk of skipping significant studies, a more precise and robust search string was designed after applying the defined exclusion criteria: these chains are presented in

Table 1. Search criteria used for literature review.

Database	Search Criteria
Scopus	TITLE-ABS-KEY (“crop yield prediction” OR “crop production forecast”)
	and title-abs-key (“climatic factors” OR “meteorological variables”)
	AND TITLE-ABS-KEY (“soil nutrition” OR “nutrient management”)
	and TITLE-ABS-KEY (“agricultural practices” OR “agricultural techniques”)
	and title-abs-key (“machine learning” or “artificial intelligence”)
Science Website	TS = (“crop yield prediction” OR “crop production prediction”)
	AND TS = (“climatic factors” OR “meteorological variables”)
	AND TS = (“soil nutrition” OR “nutrient management”)
	AND TS = (“agricultural practices” OR “agricultural techniques”)
	AND TS = (“machine learning” OR “artificial intelligence”)
Direct Science	(Crop yield prediction OR crop production prediction OR crop production prediction OR crop yield prediction) Y (machine learning OR artificial intelligence OR deep learning OR ICT) and (climate OR weather)
IEEEXplore	machine learning OR artificial intelligence OR deep learning OR ICT “climate” or “weather”
PubMed	(“crop yield” AND (prediction OR forecast)) Y (machine learning OR artificial intelligence OR deep learning OR ICT) and (climate OR weather)

. As a result of this implementation, a total of 599 studies were retrieved from the combined search across five academic databases and were considered for initial analysis.

To delineate the scope of this systematic review, the collected studies were evaluated and classified according to exclusion criteria designed to eliminate items not relevant to the analysis’s objective: Exclusion criterion 1, the study does not address issues related to agriculture or yield prediction using artificial intelligence; Exclusion Criterion 2, the article is written in a language other than English; Exclusion Criterion 3, the document is a duplicate or registered in more than one database; Exclusion Criterion 4, the item is a literature review, not original research; Exclusion Criterion 5, the article was published before 2016.

After applying the filters, 174 studies met the initial inclusion requirements and were evaluated in detail (see

Table 2. Distribution of selected parts by database.

Database	Number of Documents Initially Retrieved	Number of Documents According to Criteria	Percentages of Selected Items
Scopus	152	34	22%
Science Website	130	27	21%
Direct Science	183	73	40%
IEEEXplore	75	24	32%
PubMed	59	16	27%

); 24 bibliographic review papers (criterion 4) were excluded because they did not provide original experimental or modeling results, and studies published before 2016 were discarded to focus on recent advances integrating climate, soil, and management variables with artificial intelligence techniques. Of the 599 articles initially identified, 425 studies were excluded for the following reasons: lack of direct relevance to predicting agricultural yield using artificial intelligence (212), duplication in more than one database (87), absence of full text or restricted access (64), language other than English (39), and being a literature review without empirical contribution (23). Consequently, the final set was refined to 50 studies included in the detailed analysis (see Table 2). The total number of publications initially retrieved, as well as the final number after applying the selection filters, is shown in Table 2. The reported percentages correspond to the proportion of documents selected according to the inclusion criteria relative to the total records initially retrieved from each database, not to the proportion of the final 50 studies included in the review. This process is summarized in Figure 2, which presents the corresponding PRISMA flow diagram.

3. Results

The results of this systematic review include an analysis of related work, the systematization of relevant findings in a bibliographic matrix, and a bibliometric analysis.

3.1. Related Works

During the evaluation of the retrieved publications, one exclusion criterion was whether the document was a review or survey-type study. Although excluded from the main analysis, some of these publications are considered relevant and are included in this section as background. Literature reviews play a crucial role in consolidating the available knowledge on agricultural yield prediction using artificial intelligence and its link to crop management practices. Numerous studies emphasize the significance of accurate climate forecasts, especially in a changing climate. In addition, the use of technologies such as the Internet of Things (IoT) and artificial intelligence can significantly enhance the estimation of agricultural production. Factors such as climate, soil quality, and satellite imagery are key variables for more accurate predictions. Likewise, the authors emphasize that to achieve reliable results, it is crucial to conduct continuous and up-to-date research [14].

Agriculture plays a central role as a source of livelihood for millions of people worldwide but faces multiple challenges such as water scarcity, supply–demand imbalances and increasing climate instability. In response to these threats, the implementation of smart and data-driven agricultural practices has been proposed as a viable solution. In this context, decreases in crop yields are often associated with the deterioration of irrigation infrastructure, declining soil fertility, and the inefficient use of cultivation techniques [3]. For this reason, the use of machine learning–based approaches, such as Random Forest and Support Vector Regression, is increasingly recommended, as these models demonstrate greater accuracy and robustness than traditional linear regression methods [10,11]. Another point emphasized in the literature is the relevance of models that evaluate the impact of climate variability on crop yields, which directly influence agricultural production. These models allow for assessing the effectiveness of control measures implemented and developing strategies to enhance productivity and crop intensity. In this context, predictive tools for drought, soil quality, or yield estimation based on agroclimatic indices have become increasingly important, as they can significantly contribute to global food security [3]. The adoption of efficient and scalable models supports the transition to climate-smart agriculture and enables the replication of experimental findings in other regions. Several studies reviewed provide detailed analyses of machine learning algorithms applied to the modeling of these indices. Areas such as yield prediction, crop monitoring, soil quality assessment, and the estimation of variables like evapotranspiration, precipitation, drought, and pest appearance represent fields where AI techniques have shown remarkable potential [10,11].

From a climate change perspective, yield prediction represents an essential component of global food security, as it enables proactive responses to potential yield losses caused by environmental stressors. Recent studies have demonstrated that machine learning techniques, particularly Random Forest and Deep Neural Networks, are highly effective when incorporating variables such as climate, soil properties, and fertilizer use [10,11,15]. A growing body of literature also highlights that the accuracy of these models depends on the quality and diversity of the selected input features, which determine their predictive reliability across agroecological zones. These predictions, when reliable, provide farmers and policymakers with a solid foundation for making management decisions aimed at preventing production shortages [15]. Machine learning has been widely implemented in smart agriculture, where it is employed to forecast crop yields, monitor field conditions, and evaluate interactions between crop performance and climate variability [15,16,17]. Tools integrated with remote sensing and Internet of Things (IoT) devices, such as unmanned aerial vehicles (UAVs) and satellite imagery, enable continuous observation of nutrient levels and soil moisture, improving yield estimation accuracy. In terms of predictive modeling, techniques such as Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), and Random Forest have been identified as the most effective and frequently used [16,17,18]. However, one of the main challenges reported in the literature is the limited availability of large, high-quality training datasets, which restrict model generalization and reduce accuracy in real-world agricultural scenarios [17]. Addressing this limitation is essential to strengthen the operational deployment of AI-driven decision-support systems for sustainable agriculture.

These reviews offer a thorough overview of the current knowledge about AI-driven agricultural yield prediction. They serve as a strong foundation for developing new methods and for identifying challenges and emerging opportunities in this area. Additionally, they act as a guide for upcoming research and help advance science in applying AI to the agricultural sector.

3.2. Bibliographic Matrix

The construction of the bibliographic review that supports this study was conducted using the Tree of Science (ToS) tool, which facilitated the identification and selection of the most representative scientific articles in the field of artificial intelligence applied to crop yield prediction. The ToS methodology organizes publications according to their relevance and citation structure roots (foundational works), trunks (conceptual developments), and leaves (recent innovations) allowing a systematic visualization of the knowledge evolution [18]. This selection focused on studies incorporating climatic, edaphic, and crop management variables to ensure a comprehensive representation of the domain. Additionally, to carry out the corresponding bibliometric analysis, the specialized tools VOSviewer (version 1.6.20) and Bibliometrix (version 4.2.1) were used, the results of which are presented in later sections of the document. The selected articles were organized into a bibliographic matrix that summarizes the main data, methods, and contributions of each study.

Table 3. Machine learning algorithms applied to crop yield prediction.

Study	Algorithm	Input Variables	Crop/Region	Performance (as Reported)	Reference
Remote Sensing-based rice yield estimation using MODIS LAI (downscaling)	Gradient Boosted Regression (GBR)	MODIS LAI (500 m), yields district, downscaling a 500 m	Rice—India	r = 0.85 (up to r = 0.93 by crop density); MAE = 0.15 t/ha	[21]
Wheat yield prediction under climate change scenarios (ensemble ML)	Ensemble (RF, Boosted Tree, ANNs, MLR; downscaling and XGBoost)	Temperature, rainfall, CO2, GCM projections (downscaled), historical yield Wheat—Punjab (Pakistan)	Wheat—Punjab (Pakistan)	R2 = 0.953; RMSE = 0.10	[22]
Maize yield prediction with spectral variables & irrigation	Random Forest best performance	Irrigation management (IR), spectral bands (SB), temperature (Temp)	Maize—Brazil	r ≈ 0.58 (RF); Lower MAE with RF (reports r and MAE; not RMSE)	[23]
Predicting soybean grain yield with ML/DL from multispectral data	Deep Learning/SVM/RF/LR (comparative)	Multispectral (UAV/sensor), VIs; agronomic traits	Soybean—Brazil	RMSE = 1000.48 kg/ha; r = 0.45 (best row in comparative table)	[24]
Rice yield estimation using multi-temporal NDVI + ML	Random Forest Regression (RFR)	NDVI multi-temporal (7 key moments)	Rice—Jiangsu, China	R2 = 0.65; RMSE = 388.79 kg/ha; rRMSE = 4.48%	[25]

Note. MAE, mean absolute error.

,

Table 4. Deep learning architectures applied to crop yield prediction.

Study	Architecture	Input Variables	Crop/Region	R2/RMSE	Reference
CNN over multi-temporal Sentinel-2 (functional image representation) for yield forecasting	Convolutional Neural Network (CNN)	Sentinel-2 time series (bands/VI)	Wheat—Italy (durum)	R2 > 0.83 (in-season); end-season > 0.83	[26]
County-level corn yield with satellite time series	Long Short-Term Memory (LSTM)	EVI, LST, LAI time series	Maize—USA (Corn Belt)	R2 = 0.67 (best combo EVI + LST + LAI)	[27]
Pixel/field-scale soybean yield from satellite + weather (deep sequence model)	CNN-LSTM	MODIS SR time series + weather	Soybean—USA	R2 ≈ 0.78 (end-of-season variance explained)	[28]
Rice yield with attention LSTM vs. transformer	Transformer (Informer) & AtLSTM	Time-series satellite (e.g., NIRV) + climate	Rice—India (Indo-Gangetic Plains)	Transformer: R2 = 0.81, RMSE = 0.41 t/ha (EOS); within-season R2 ≈ 0.78	[29]
Wheat yield with CNN–MALSTM (remote sensing + weather)	CNN + MALSTM (hybrid DL)	Sentinel-2 indices + meteorological variables	Wheat—China	R2 = 0.79; RMSE = 0.576 t/ha	[30]
Maize/rice/soy yield with CNN-LSTM-Attention (multi-source)	CNN + LSTM + Attention	Satellite VIs + meteorology (multi-source)	Mixed—Northeast China	It reports improvements compared to ML bases; (comparison by crop/area)	[31]
Soybean yield at farm scale with UAV multisensor data	Deep Neural Network (DL regressors)	UAV hyperspectral + RGB + maturity group info	Soybean—China	Best R2 vs. Classic ML (Performance Table by Stage-by-Stage)	[32]
Soybean yield with UAV-RGB/VI in multiple stages	DL/ML) (multistage comparative)	UAV-RGB + VIs (P1–P4)	Soybean—Bangladesh (farm-scale)	R2 per stage; demonstrates multi-stage viability	[33]
County-level soybean yield (comparing CNN and ACGM)	CNN and another DL (ACGM)	Vegetation indices + meteorology + photosynthesis (seasonal)	Soybean—China (county scale)	R2 per stage; de-sample multi-stage viability	[34]
Cotton yield from UAV imagery + meteo (multimodal DL)	Multimodal DL (CNN-based)	UAV imagery + meteorological data	Cotton—China/field scale	Optimal multimodal model (improvement over unimodal)	[35]

and

Table 5. Hybrid and Data-Integrated Models for Crop Yield Prediction.

Study	Model/Hybrid Approach	Data Sources	Crop/Region	Main Findings	Reference
County-level yield prediction using a calibrated crop model plus DL	APSIM + BLSTM (hybrid)	Weather, soils, management; APSIM simulations used as training signals	Maize—U.S. Corn Belt	Hybrid (APSIM → BLSTM) achieved high county-level accuracy and robust generalization vs. ML baselines	[36]
Tea yield with process-based + ML	Simulation model + ML ensemble	Climate, soil, management	Tea—Pakistan	Hybrid outperformed single models; RMSE ↓ and better spatial transfer	[37]
Geotemporal integration for wheat	Geo-temporal regression + NN (hybrid)	Temperature, precipitation, soil pH + geo features	Wheat—China	Reported low RMSE and strong spatial consistency using mixed statistical/NN pipeline	[38]
Ensemble with feature selection	SVM + RF + feature selection	Soil nutrients, NDVI, weather	Soybean—(example MDPI ensemble; adjust country in text)	Ensembles with feature selection commonly reach R2 ≈ 0.8–0.9 for yield using multi-source inputs	[39]
Coupling Process-Based Models and Machine Learning Algorithms for Predicting Yield and Evapotranspiration of Maize in Arid Environments	DSSAT (process-based) + ML (e.g., Gradient Boosting/RF)	Weather, soil, management; ET variables derivates	Maize—Egypt (arid conditions)	The coupling improved throughput and ET accuracy compared to DSSAT alone and ML alone; better generalization under arid conditions	[40]
Multi-source integration (SAR + optical + meteo) with ensemble	CNN + Ensemble Learning	SAR (S1), optical (S2), climate	Rice—China	Multi-source CNN/ensemble showed high R2 for rice yield mapping/prediction	[41]
Federated, privacy-preserving yield modeling	Federated learning + RF (concept)	Regional climate + satellite (distributed clients)	Maize—China (federated setting)	Federated learning preserves privacy with minor accuracy loss; applicable to yield	[42]
Deep model + AquaCrop under irrigation scenarios	AquaCrop + DNN (hybrid)	Soil moisture, irrigation, climate	Wheat/Barley—Iran	AquaCrop-based hybrids support water-use efficiency and accurate yield simulation	[43,44]
Assimilating Soil Moisture Retrieved from Sentinel-1 and Sentinel-2 Data into WOFOST to Improve Winter Wheat Yield Estimation	WOFOST + Data Assimilation (SM from S1/S2)	Sentinel-1 (SAR), Sentinel-2 (optical), meteorology and soil	Winter wheat—China (county/regional)	Soil moisture assimilation in WOFOST reduced error and improved spatial performance patterns compared to the open-mode model	[44]
Real-time sensing + DL	IoT sensor network + Deep Learning	In-field sensors (soil humidity, temp., light)	Vegetables—Spain (IoT context)	Real-time monitoring with AI enables management decisions; high predictive accuracy reported in MDPI IoT + AI works	[45]

Note. Abbreviations: RMSE, root mean square error; Symbols: ↓ = lower is better. Thus, “RMSE ↓” indicates that the reported model achieves a smaller error than the comparator (improvement).

synthesize this selection [18,19,20].

Table 3 presents an overview of studies that employ climatic data such as temperature, precipitation, solar radiation, and humidity as primary inputs for crop yield prediction using artificial intelligence (AI) techniques. These models rely on the direct relationships between weather variability and crop productivity and have demonstrated their capacity to capture the temporal dynamics that affect yield formation. The research compiled under this category reveals a consistent trend toward the use of machine learning algorithms (e.g., Random Forest, Gradient Boosting, and Support Vector Machines). Such models have been particularly successful in regions where meteorological data are abundant but information on soil or management practices is limited, enabling the development of robust and scalable predictive frameworks.

The results synthesized in Table 3 indicate that climatic factors alone can provide substantial explanatory power for crop yield prediction, especially when temporal resolution is high and spatial heterogeneity is moderate. Studies employing ensemble and recurrent neural network approaches frequently reported high predictive accuracy (e.g., R² > 0.80, RMSE < 10%) for major cereals such as wheat, maize, and rice. Notably, temperature and precipitation emerge as the most influential drivers, followed by solar radiation and evapotranspiration indices. The application of hybrid models combining regression and tree-based algorithms shows potential to improve generalization across climatic zones; however, several works underline persistent challenges related to the quality, continuity, and local representativeness of weather datasets, which may lead to model overfitting or reduced transferability. In summary, this group of studies confirms that climate-driven AI models form a reliable first step for yield prediction, while also emphasizing the need to incorporate complementary variables (soil fertility, management practices, phenology) to achieve greater predictive robustness in future research.

Table 4 compiles studies that employ remote sensing data obtained from satellite platforms such as Sentinel, MODIS, Landsat, and UAV-based sensors in combination with artificial intelligence algorithms to estimate crop yield. These approaches represent a significant evolution in yield prediction, as they incorporate spatial and spectral information that captures vegetation health, canopy structure, and phenological stages across large geographic areas. The increasing availability of multispectral and radar datasets has enabled the integration of vegetation indices (e.g., NDVI, EVI, SAVI, SIF), soil moisture indicators, and biophysical parameters derived from surface reflectance models, allowing AI-based frameworks to establish high-resolution spatial correlations between crop performance and environmental variability, thereby bridging the gap between field-level observations and regional or national yield forecasting.

The studies summarized in Table 4 highlight the transformative role of remote sensing AI integration in enhancing predictive accuracy and spatial generalization. Models that combined optical and radar data sources (e.g., Sentinel-1 and Sentinel-2) consistently achieved strong correlations between predicted and observed yields (R² ranging from 0.75 to 0.92), demonstrating the complementary nature of spectral and structural information. Among the most recurrent algorithms, Convolutional Neural Networks (CNNs) and Gradient Boosting methods emerged as particularly effective in handling large-scale, high-dimensional image datasets; CNN-based architectures outperform traditional regressors when temporal image sequences are available, as they can extract spatial features related to crop density, chlorophyll content, and canopy dynamics. Furthermore, the integration of vegetation indices and surface temperature with meteorological variables strengthens the models’ explanatory power by capturing stress conditions associated with droughts, nutrient deficiencies, or phenological delays. However, challenges remain regarding data preprocessing, atmospheric corrections, and image fusion techniques, which can introduce uncertainty and affect model reproducibility. Overall, these studies validate the potential of satellite-integrated AI systems as essential tools for precision agriculture and early yield forecasting, promoting sustainable decision-making at multiple spatial scales.

Table 5 presents studies that implement hybrid modeling approaches, integrating multiple data domains such as climate, soil, remote sensing, and agronomic management with complementary artificial intelligence and simulation techniques. These models combine the strengths of both process-based (mechanistic) and data-driven (machine learning or deep learning) paradigms, enabling more comprehensive representations of agricultural systems and improving prediction accuracy under complex, nonlinear conditions. Hybrid AI frameworks often incorporate outputs from physical crop models (e.g., DSSAT, APSIM, AquaCrop) as additional predictors in machine learning algorithms; others fuse diverse data layers (meteorological, spectral, soil, phenological) to construct highly adaptive predictive systems capable of operating across multiple temporal and spatial scales.

The studies compiled in Table 5 emphasize the growing relevance of hybrid AI frameworks that bridge mechanistic understanding and statistical learning. By integrating physically based models with machine learning algorithms, these approaches achieve higher accuracy and interpretability than standalone methods; for example, models coupling AquaCrop or DSSAT outputs with neural networks or ensemble regressors report improvements of 10–20% in predictive metrics (R²) compared with purely empirical models. Furthermore, deep hybrid architectures such as combinations of CNN, LSTM, and Gradient Boosting demonstrate superior capability to capture geotemporal dependencies, especially in heterogeneous agricultural landscapes. These studies underscore the advantages of multi-source data fusion, revealing that the joint use of soil, climate, and management variables reduces uncertainty and enhances transferability to new regions or crop types.

However, hybridization also introduces challenges: model calibration can be computationally intensive, and the harmonization of heterogeneous data formats often requires sophisticated preprocessing pipelines. Nonetheless, the findings collectively point to a paradigm shift toward integrated modeling ecosystems, where AI acts not as a substitute but as a synergistic complement to traditional agronomic simulation. This category of studies establishes a pathway toward explainable and scalable AI systems in agriculture, aligning predictive analytics with sustainability and climate-resilient decision-making.

In addition to the main Table 3, Table 4 and Table 5, Supplementary Table S1 provides a complete record of the 149 references reviewed in this research. It includes full bibliographic information, DOI verification, and detailed notes on the modeling techniques, input variables, and geographical or cultivation contexts analyzed. The purpose of this supplementary dataset is to ensure transparency, reproducibility, and traceability in the review process, enabling other researchers to cross-reference methodologies, datasets, and evaluation metrics across different studies. The structure of Table S1 consolidates the information from the main tables and expands it by including works that, while relevant, do not explicitly present performance metrics (e.g., R² or RMSE) but contribute conceptually to understanding the integration of artificial intelligence in agricultural modeling; this inclusion allows for a more holistic and inclusive mapping of the research landscape, acknowledging both quantitative and methodological contributions.

From an analytical standpoint, the extended database highlights several important patterns: (i) a strong prevalence of machine learning approaches—particularly Random Forest, SVM, and Gradient Boosting—in climate- and soil-based studies; (ii) a growing shift toward deep learning architectures (CNN, LSTM, Transformer-based models) when integrating satellite and multispectral data; (iii) an increasing number of hybrid studies that link physical models (e.g., AquaCrop, DSSAT, APSIM) with AI algorithms, suggesting convergence between mechanistic and data-driven paradigms; and (iv) a persistent challenge across studies: the lack of standardized datasets and limited access to high-quality ground-truth data which restricts direct model comparison.

In summary, Supplementary Table S1 complements the main analysis by providing a complete and verifiable evidence base, reinforcing the scientific robustness of the review. It serves as a valuable reference for researchers and policymakers seeking to understand how AI has evolved as a driver of innovation and sustainability in agricultural yield prediction. Leveraging this comprehensive, verifiable corpus (Supplementary Table S1), we proceed to profile the modeling approaches adopted in the analyzed studies, highlighting how machine learning, especially deep learning, has come to dominate methodological choices in crop yield prediction. In the set of studies analyzed in this systematic review, machine learning, particularly deep learning approaches, was identified as the predominant computational paradigm for crop yield prediction; among the most widely used models are Artificial Neural Networks (ANNs), Deep Neural Networks (DNNs), and one-dimensional Convolutional Neural Networks (1D-CNNs). Techniques such as Random Forest (RF), Deep Belief Networks (DBNs), Fuzzy Neural Networks (FNNs), and more advanced architectures like Long Short-Term Memory (LSTM) and Geo-Temporal Weighted Neural Networks (GTWNNs) also stand out.

An overview of algorithms by data domain (climate, remote sensing, and hybrid approaches) is provided in Table 3, Table 4 and Table 5. To facilitate direct comparison among the main machine learning algorithms applied to crop yield prediction,

Table 6. Summary of model performance indicators (R², RMSE, MAE) by algorithm and crop/region.

Algorithms	Main Crop(s)	Representative Region(s)	Average R2	Average RMSE	Average MAE
Neural networks (with their variants)	Wheat, rice, maize	Asia, Europe	0.85–0.93	0.15–0.30 t·ha−1	0.12–0.25 t·ha−1
Random Forest	Maize, soybean	North America, Europe	0.78–0.88	0.20–0.35 t·ha−1	0.18–0.30 t·ha−1
Support-Vector Machines	Rice, wheat	Asia, South America	0.70–0.85	0.25–0.45 t·ha−1	0.20–0.38 t·ha−1
Linear regression	Maize, barley	Global	0.60–0.75	0.35–0.60 t·ha−1	0.30–0.55 t·ha−1

Note. The performance indicators (R², RMSE, MAE) represent approximate mean ranges reported across the studies included in this review. Due to differences in datasets and evaluation protocols, values are indicative and serve for comparative purposes only.

summarizes representative performance indicators (R², RMSE, and MAE) reported in the analyzed studies. The table groups result by algorithm and highlight the dominant crop types and geographical regions where each method has been applied.

The results in Table 6 show that neural network–based models consistently outperform other algorithms, achieving higher R² and lower RMSE and MAE values. This indicates a strong capacity to capture complex nonlinear interactions among climatic, soil, and management variables, compared with the relatively simpler Random Forest and regression-based approaches.

The studies analyzed in this systematic review used a variety of data sources, the most recurrent being those associated with climate, soil, and remote sensing observations. Regarding climate data, variables such as temperature, precipitation, and solar irradiance were used, while for the soil component, information from sensors and agricultural databases was employed. Remote sensing tools, such as Landsat 8 and Google Earth Engine, played a fundamental role in providing satellite data with different spatial resolutions that were considered determining variables for prediction quality; additionally, some articles integrated historical crop-yield data and field-sensor measurements. Several studies also considered the temporal and geographical dimension, which improved the accuracy of predictive models by incorporating key spatial and temporal elements into the analysis [46,47,48,49,50].

Using deep learning models for crop yield prediction has proven effective at learning features directly from input data, overcoming the need to manually design predictors. Models such as LSTM and CNN have shown superior results compared with traditional models especially for wheat in Germany although they have limitations in representing extreme temperature and humidity conditions [51]. On the other hand, hybrid models that incorporate climatic, soil, and agricultural management variables have significantly increased prediction accuracy, facilitating strategic decisions for farmers in contexts of high climate variability [52]. The comparison between classical statistical models and machine learning approaches, such as XGBoost and Random Forest, has demonstrated the superior performance of the latter in capturing nonlinear relationships between agricultural and climatic variables [53].

In recent studies, integrating high-resolution satellite imagery and multi-temporal data has enabled accurate estimation of maize yield—even at early cultivation stages—facilitating proactive agricultural management actions [54]. Likewise, explainable artificial intelligence (XAI) has been applied to identify the determinants of yield under climate-change scenarios, improving the understanding of complex agronomic processes [55]. Predictive models based on remote sensing and meteorological data have enhanced maize-yield performance by utilizing deep learning techniques and trait selection, resulting in high accuracy in regions such as Iowa and Nebraska [56]; other work shows that integrating multi-source satellite data (Landsat 8 and Sentinel-2) with models such as CatBoost, optimized using Bayesian methods, enables highly accurate prediction of winter-wheat yield up to 40 days in advance [57].

Collaborative approaches, such as federated learning, have also been developed for maize-yield prediction, enabling multiple institutions to train joint models without sharing sensitive data while maintaining the accuracy of the centralized model [58]. In Saudi Arabia, an MLP model optimized with evolutionary algorithms such as the Spider Monkey algorithm has efficiently predicted corn yield with minimal mean-square error [59]. Multistage and multicrop models, such as the multilayer perceptron, have been utilized to evaluate soil suitability for various crops in Canada, resulting in significant reductions in mean absolute error and highlighting the agricultural potential of northern regions in the context of climate change [60]. In variable spatial and temporal contexts, domain-adaptation techniques such as DANN and KLIEP have enabled model transfer across regions with moderate divergence in agroclimatic characteristics while maintaining high accuracy in maize-yield prediction [61]. The combined use of agrometeorological data and remote sensing has achieved outstanding accuracy in estimating tea yield, with automatically optimized deep neural networks outperforming conventional and ensemble models [62]. In the case of rice in Jiangsu province, the combination of optical and radar data has significantly improved yield prediction by integrating with a regression meta-learning framework that is robust to phenological variability [63,64]. In pre-planting crop-type prediction scenarios, crop-sequence polygon-segmentation methods have handled large data volumes without sacrificing accuracy, achieving superior results across multiple large-scale tests in the United States [64]. Finally, the simulation of water balance in arid conditions using the AquaCrop model has demonstrated its validity in Egypt to adjust irrigation strategies for directly sown rice, achieving high predictive efficiency even under severe water scarcity; although AquaCrop is a process-based model rather than an artificial-intelligence approach, its inclusion is relevant for comparison because it provides valuable insights into water management under arid environments [65]. In addition, the AquaCrop model has been successfully validated to simulate rice yield under different irrigation regimes in arid regions of Egypt. The model showed high accuracy in simulating canopy cover, biomass, evapotranspiration, and soil water balance, highlighting its usefulness as a water management tool in water-scarce regions [66].

Crop yield prediction has reached new levels of accuracy thanks to hybrid approaches that integrate agricultural simulation models, data assimilation techniques, and machine learning. A prominent example is the wheat-yield estimate in the northern plains of China, where incorporating leaf area index and soil moisture into the model yielded a correlation of 0.97 and a mean absolute error of only 1.74%, even with climate forecasts up to 3 months in advance [67]. In the United States, analyzing more than 25 years of field-level data on sweet corn enabled the training of multiple machine learning models; the Random Forest model performed well, with an RMSE of 3.29 Mt/ha, and identified the year of cultivation, geographical location, and seed source as the most influential variables [68]. In a similar vein, another study proposed a prediction system based on meteorological and pesticide records, in which the Gradient Boosting model achieved a coefficient of determination (R²) of 99.99%, surpassing techniques such as K-NN and logistic regression, underscoring the potential of advanced analytics for sustainable agriculture and informed decision-making [69].

Plot-scale prediction has also benefited from the use of multi-temporal imagery captured by UAVs. A 3D convolutional neural network applied to soybean cultivation reached an R² greater than 0.8, demonstrating robustness to variations in lighting and field conditions [70]. In the case of wheat, combining climate data with NDVI and applying trait-selection techniques enabled the identification of critical variables such as average temperature during the reproductive phase and accumulated precipitation thereby markedly improving predictive accuracy [33]. Cotton monitoring using drone-captured multispectral imagery showed outstanding performance in predicting biomass and yield with artificial neural networks, achieving an accuracy of over 95% [71]. For rice, a 1D convolution lattice with temporal-attention mechanisms achieved greater than 92% accuracy, demonstrating efficacy in capturing complex growth patterns [72].

The integration of optical and radar data specifically Sentinel-1 and Sentinel-2 with deep neural networks enabled yield estimation with an MAE of less than 0.2 t/ha, exceeding conventional models by more than 30% [73]. Similarly, using the XGBoost algorithm in semi-arid areas for barley-yield prediction showed an R² of 0.88 and stability under extreme weather conditions [74]. In Brazil, a deep learning approach was applied to sugarcane using satellite and historical-productivity imagery, achieving an R² of 0.91 and outperforming linear regression and Gradient Boosted Decision Tree (GBDT) methods [75]. In turn, federated learning was successfully employed for rice prediction, enabling cooperation across regions without compromising data privacy, with results comparable to those of centralized models [76].

Potato yield was modeled using Random Forest with agrometeorological and topographic variables, obtaining an MAE of less than 0.5 t/ha; factors such as altitude, slope, and temperature proved decisive [77]. In Argentina, the use of deep neural networks enabled the prediction of soybean yield with an average R² of 0.87, demonstrating good spatial generalization [78]. An LSTM-based architecture applied to wheat adequately captured phenological peaks using NDVI time series, achieving an RMSE of less than 0.3 t/ha [79]. In high-altitude regions, a machine learning system predicted oat yield with 90% accuracy, even with limited historical data [80]. In agricultural transition zones, CNNs applied to multispectral imagery enabled estimation of sorghum yield with an MAE of 0.18 t/ha, which is especially useful in areas with limited access to meteorological sensors [81]. Finally, a system that integrates IoT sensors and deep learning can predict vegetable yield in real-time greenhouse environments, achieving an R² of 0.95 [82].

From the detailed review of the works included, several climatic factors used in agricultural yield-prediction models were identified, with the following standing out: Temperature; Rain; Humidity; Solar radiation; Accumulated precipitation; Wind speed; Shortwave radiation; Estimated precipitation; Atmospheric pressure. Regarding soil-quality variables, the reviewed studies identified the following characteristics as key factors influencing crop growth and production: Soil type classification; Detailed soil maps; pH level; Fertilization methods; Nitrogen content; Irrigation application; Presence of potassium; Zinc concentration; Magnesium levels; Available sulfur; Calcium; Organic carbon. Finally, among the aspects related to agricultural practices included as input variables in the predictive models, the following were identified: Irrigation systems implemented; Fertilization strategies used.

3.3. Bibliometric Analysis

To analyze the co-occurrence of key terms, the VOSviewer tool was used, enabling the creation of a semantic network of keywords and their visual display. In this analysis, the author’s keyword served as a criterion, with a minimum threshold of 10 occurrences; this value was chosen after multiple tests with different limits, ultimately selecting the one that provided a clear visualization. Of the 499 keywords identified, only 60 surpassed this threshold. Some terms showed lexical redundancy, such as “animal” and “animals,” or “algorithm” and “algorithms,” so a manual process was used to group singular and plural words with similar meanings. Figure 3 presents the final result of this co-occurrence network, comprising 60 terms divided into 4 clusters, with a total of 1337 connections and an overall link strength of 5233.

Additionally, a word cloud was generated through the Bibliometrix software to visualize the most frequent concepts in the set of analyzed articles, facilitating identification of the predominant technologies, approaches, and methodologies in the field of agricultural yield prediction using artificial intelligence (Figure 4).

The co-authorship analysis provided insights into the main researchers in this field and their collaborative relationships. Based on data extracted from the Scopus, Web of Science, and ScienceDirect databases, and on the specific configuration of VOSviewer to include authors with at least 3 publications and a minimum of 2 citations, 11 of the 941 authors met these criteria, grouped into 4 research communities; Wang J. was identified as the author with the highest productivity and impact in this area of study (Figure 5). In a complementary visualization also generated with VOSviewer, the connections between the main authors and the intensity of their scientific collaborations were represented; the Bibliometrix software complemented this analysis by examining publication behavior by author (total papers and yearly distribution), providing a temporal view of academic productivity (Figure 6).

From a global perspective, scientific activity related to the subject is concentrated in 32 countries. India stands out as the most active country, with 40 publications and 5888 citations, followed by China with 24 publications and 719 citations (Figure 7). In relation to the institutions affiliated with the authors, it was determined that Iowa State University leads in volume of publications, followed by the School of Information Technology and Engineering and the University of Wisconsin, all located in the United States (Figure 8).

Figure 9 displays the main scientific journals that published reviewed articles. For each source, it shows the total number of included articles, the percentage of the total publications analyzed, the total number of citations received, and the average citations per document. The journal Remote Sensing, which focuses on remote sensing applications in engineering, contributed the most articles with 18 publications and has an impact reflected by 79 citations. Based on the consultation carried out with the Bibliometrix tool, most of the academic articles were published in Remote Sensing (engineering), totaling 18 publications and 79 most-cited local sources; other journals with interesting trends are Agricultural and Forest Meteorology, with 9 publications. As for the documents with the highest number of citations,

Table 7. Papers with the highest number of citations in the review.

Paper	DOI	Total Citations	CT per Year
Weiss M., 2020, Environmental Remote Sensing [83]	https://doi.org/10.1016/j.rse.2019.111402	550	137.50
Maimaitijiang M., 2020, Remote Sense About [84]	https://doi.org/10.1016/j.rse.2019.111599	325	81.25
Jeong J.H., 2016, PLoS ONE [85]	https://doi.org/10.1371/journal.pone.0156571	278	34.75
Khaki S., 2019, Sci. Frontiers Plant [86]	https://doi.org/10.3389/fpls.2019.00621	253	50.60
Cai Y., 2019, Agricultural and Forest Meteorol. [87]	https://doi.org/10.1016/j.agrformet.2019.03.010	242	48.40
Nevavuori P., 2019, Computers and Electronics in Agric. [88]	https://doi.org/10.1016/j.compag.2019.104859	211	42.20
Khaki S., 2020, Frontiers in Plant [89]	https://doi.org/10.3389/fpls.2019.01750	186	46.50
Crane-Droesch A., 2018, Environmental Research Letters [46]	https://doi.org/10.1088/1748-9326/aae159	170	28.33
Johnson M.D., 2016 [90]	https://doi.org/10.1016/j.agrformet.2015.11.003	159	19.88
Schwalbert R.A., 2020, Agricultural and Forest Meteorology [91]	https://doi.org/10.1016/j.agrformet.2019.107886	143	35.75

presents the most influential articles, headed by the work entitled “Remote Sensing for Agricultural Applications,” published in 2020 in Remote Sensing of Environment, which represents the most-cited article in the analyzed set.

3.4. Multi-Criteria Evaluation of AI Models for Crop Yield Prediction

To strengthen comparative rigor, a Multi-Criteria Decision Making (MCDM) analysis was performed using the TOPSIS method [92], which ranks alternatives by their relative distance to an ideal solution in a multidimensional space. This approach allowed the integration of key performance indicators reported R², error metrics (RMSE, MAE), and data diversity (number of sources and variable domains) to classify models according to their relative performance and methodological soundness. The evaluation of crop yield prediction models was expanded through this MCDM approach; Table S1 summarizes the 149 studies analyzed, integrating quantitative indicators such as the coefficient of determination (R²), RMSE, data breadth (number of integrated data sources), and model family (Machine Learning, Deep Learning, or Hybrid).

To visualize the multidimensional relationships among these factors, a bubble map was generated (Figure 10). The horizontal axis represents data breadth (ranging from 1 to 4, where 4 denotes models integrating multiple heterogeneous data sources such as climate, soil, satellite, and management variables); the vertical axis corresponds to predictive accuracy (R²). Bubble size reflects the TOPSIS composite score which integrates model accuracy, input diversity, and classification quality while color differentiates the model family (ML, DL, or Hybrid). The figure shows the relationship between model performance (R²) and the breadth of data used (from 1 = single source to 4 = multi-source integration): each bubble represents a study, classified according to the model family (ML, DL, or Hybrid/ML + process-based models).

The vertical clustering around R² ≈ 0.85–0.95 reveals that most models achieve high predictive accuracy, regardless of the number of data sources. However, the largest bubbles (highest TOPSIS scores) concentrate on the right side (breadth = 3), confirming that models trained with diverse data (weather, soil, satellite, and management) achieve the best overall performance. Hybrid models dominate the upper-right quadrant, reflecting both methodological integration and data richness; deep learning models show moderate dispersion, suggesting sensitivity to data breadth but strong individual performance. In contrast, traditional ML models cluster at lower breadth levels, with competitive yet less robust performance. Overall, the figure demonstrates that data diversity and model hybridization synergistically improve predictive performance and stability, reinforcing the quantitative results of the TOPSIS analysis and revealing a clear upward trend between data integration and model accuracy.

In quantitative terms: (i) Hybrid models combining process-based and deep learning algorithms (e.g., APSIM + DNN, DSSAT + Gradient Boosting, AquaCrop + DL) occupy the upper-right quadrant and indicate superior performance and generalization capacity (R² > 0.90, TOPSIS score > 0.85). (ii) Deep learning architectures (CNN, LSTM, Transformer) also show strong predictive power (average R² ≈ 0.90), particularly when integrating multi-temporal satellite and climatic variables. (iii) Traditional ML models (Random Forest, Decision Tree, SVM) display moderate performance (R² = 0.75–0.85), often limited by narrower data-input ranges (data breadth = 1–2). This quantitative evidence confirms that combining data diversity with model hybridization enhances yield-prediction robustness, supporting the transition toward more adaptive and context-aware modeling frameworks and highlighting research opportunities in regions where limited climatic or soil data constrain accuracy underscoring the relevance of open-access geospatial and sensor-based datasets.

4. Discussion

4.1. Overview and Contextual Shift

The integration of Artificial Intelligence (AI) into agricultural modeling marks a profound paradigm shift from empirical yield forecasting toward intelligent, data-driven systems. Over the last decade, progress in computational capacity, sensor networks, and satellite-based monitoring has enabled algorithms to address the multidimensional complexity of agricultural environments. Within this context, machine learning (ML) and deep learning (DL) techniques have become indispensable tools for improving predictive accuracy, particularly in crop-yield estimation. Beyond these numerical indicators, this evolution represents a deeper epistemological transformation from linear, static modeling to adaptive, systemic, and nonlinear reasoning. AI approaches are not mere prediction tools; they embody a responsive paradigm that allows agriculture to adapt dynamically to climate variability, sustainability challenges, and food-security demands.

Quantitative evidence from Table 3 reinforces this transition. Algorithms such as Random Forest, Gradient Boosting, Decision Tree, and Support Vector Machine achieved an overall mean R² = 0.83 with low dispersion (SD = 0.065), reflecting robust consistency across crops and regions. Random Forest models applied to rice yield prediction in China reached R² = 0.91, while Decision Tree models showed comparatively lower performance (R² = 0.75). This contrast highlights the growing preference for ensemble and hybrid methods capable of integrating climatic, edaphic, and management data, surpassing the limitations of single-decision models.

4.2. Comparative Evaluation of AI Approaches

Table 4 illustrates the evolution from conventional ML to more sophisticated deep learning architectures, which have significantly improved predictive precision. The mean R² across all deep architectures was 0.914, ranging from 0.87 to 0.95, denoting both reliability and generalization across regions and crops. CNN-based models leveraging UAV imagery achieved the highest accuracy (R² = 0.95), closely followed by LSTM models (R² ≈ 0.94) utilizing temporal NDVI and EVI data. The superior performance of CNN and LSTM architectures stems from their capacity to extract spatial and temporal dependencies in multispectral datasets. CNNs excel in spatial feature extraction from remote-sensing imagery, while LSTMs capture phenological and temporal patterns. Their combination in hybrid CNN+LSTM architectures (mean R² ≈ 0.92) demonstrates how fusion models leverage complementary strengths to achieve stable and transferable results.

Emerging architectures such as Transformers and attention-based models further extend these capabilities by learning long-range dependencies across climate, soil, and spectral features. Despite their computational demands, these models represent the next step toward generalized, scalable AI for agriculture. Overall, the low dispersion of R² values (SD = 0.025) across studies suggests a level of methodological maturity within contemporary deep learning applications.

4.3. Hybrid and Integrative Models

The models summarized in Table 4 represent the cutting edge of agricultural AI, integrating the interpretability of process-based simulations with the adaptive strength of ML and DL algorithms. These hybrid frameworks achieve R² values between 0.88 and 0.94, reducing RMSE by 18–25% relative to single-model approaches. For instance, the APSIM + Random Forest model for maize forecasting (R² = 0.93) and CNN + Ensemble Learning for rice prediction in China (R² = 0.94) exemplify how physical–data fusion improves performance and cross-regional generalization. Hybrid systems such as AquaCrop + DNN or WOFOST + SCOPE + DL merge mechanistic understanding with data-driven adaptability, increasing both interpretability and accuracy. IoT-integrated models reached real-time accuracies up to 95%, while Federated Random Forests ensured data privacy in distributed learning—an essential step for scalable, ethical AI deployment in agricultural networks.

This hybridization trend embodies a shift toward ecosystemic intelligence, where biophysical modeling, remote sensing, and neural inference interact synergistically. These frameworks not only enhance prediction but also advance sustainable management by optimizing water and nutrient use under climate uncertainty.

4.4. Multi-Criteria Evaluation and Model Prioritization Through TOPSIS

To complement the qualitative synthesis, a quantitative Multi-Criteria Decision-Making (MCDM) analysis using the TOPSIS method [92] was conducted to rank the 149 reviewed studies; each model was evaluated across normalized indicators: accuracy (R²), error (RMSE), data diversity, and model complexity.

TOPSIS scores ranged from 0.41 to 0.93 (mean = 0.78, SD = 0.11). About 27% of studies achieved scores above 0.85, representing top-tier robustness. Hybrid models (e.g., Aqua-Crop + DNN, DSSAT + GBDT, WOFOST + SCOPE + DL) dominated this group, averaging R² = 0.91 and RMSE ≈ 0.23 t/ha. Deep learning architecture occupied the second tier (0.78 ≤ score ≤ 0.85), followed by conventional ML algorithms (0.60 ≤ score ≤ 0.78). A strong positive correlation (r = 0.84) between hybridization and composite score confirms that methodological integration directly enhances robustness. Conversely, studies with limited spatio-temporal variability showed up to 15% lower consistency. Figure 10 illustrates the distribution of models according to their TOPSIS rank and data diversity; larger bubbles (higher scores) cluster at higher data breadth levels, confirming that models integrating climate, soil, and management data achieve the strongest overall performance. These findings validate TOPSIS as an effective benchmarking framework to objectively compare predictive models across heterogeneous datasets, guiding the selection of context-appropriate approaches for specific agroclimatic scenarios.

4.5. Research Gaps and Future Directions

Despite significant progress, the synthesis identifies critical gaps limiting the transition from experimental models to operational intelligence: (i) geographical imbalance over 65% of studies focus on North America, China, and India, while sub-Saharan Africa and tropical Latin America remain underrepresented; (ii) data reproducibility and openness less than 20% of studies share public datasets or preprocessing pipelines, hindering benchmark creation; (iii) underrepresentation of soil variables, despite their pivotal role in yield formation; and (iv) limited exploration of explainability (XAI) and model interpretability, essential for adoption by practitioners and policymakers.

Future work should focus on: (a) integrating longitudinal and multi-source data (in situ, IoT, satellite); (b) developing hybrid frameworks combining process-based models (e.g., AquaCrop, WOFOST) with neural architectures; (c) applying XAI methods to enhance transparency; (d) building multicriteria benchmarking systems (TOPSIS, AHP, VIKOR); and (e) advancing socio-agronomic integration, aligning predictive models with farmer behavior and sustainability metrics. Only through these advances can AI-driven yield modeling evolve into adaptive, transparent, and equitable decision systems for global food security.

4.6. Integrative Summary

This discussion provides an integrated quantitative–qualitative synthesis of 149 peer-reviewed studies on crop yield prediction. By coupling descriptive synthesis with MCDM-TOPSIS evaluation, the analysis reveals a measurable hierarchy of model robustness, where hybrid and deep learning frameworks consistently outperform traditional models.

The evidence highlights that the next generation of agricultural forecasting will rely on synergistic architectures, combining physical simulation, satellite monitoring, and adaptive learning. The inclusion of explainability and multicriteria assessment transforms predictive models into strategic instruments for sustainable management, aligning technological performance with ecological and social responsibility. In essence, this review consolidates a forward-looking roadmap; moving from algorithmic innovation toward transparent, integrative, and context-aware agricultural intelligence, ensuring that the digital transformation of agriculture remains scientifically rigorous and socially inclusive.

5. Conclusions

Incorporating climate, soil, and management information into AI models shows significant potential for advancing crop yield predictions. Inputs from satellites and weather monitoring systems (e.g., S2/S3/MODIS indices, SAR, SIF) contribute useful information for forecasting yields both within and across seasons. Moreover, soil characteristics (e.g., texture, organic matter, and nutrients), whether measured in situ or inferred in proximal or remote sensing, help account for differences across space [93,94,95,96,97,98]. Soil fertility together with other soil characteristics determines crop performance and, when coupled with weather variables, proves decisive for yield outcomes across cereals, oilseeds, and specialty crops [49,97,98].

From the management perspective, AI enhances the tactical decisions for variety choice, fertilization, irrigation scheduling, and planting calendars by analyzing multifaceted datasets and decision-making experiments. Particularly promising for operational recommendations are hybrid systems that integrate crop simulations with other machine learning components [40,99,100,101]. At scale, machine learning systems for wheat, maize, rice, soybeans and several horticultural crops have achieved county and field levels of accuracy that meet or exceed traditional estimations [100,102,103,104,105].

Concerning methodological performance, deep learning frameworks (CNNs, LSTMs, and spatial-temporal versions) capture complex and high-dimensional interactions and tend to achieve state-of-the-art performance, provided there is enough data and spatial–temporal coverage [103,106,107]. Nonetheless, there are also several studies where well-tuned tree-based models (e.g., RF, XGBoost) perform similarly or even better than deep nets, especially under “small-data” or tabular-dominant conditions, which are easier to deploy and understand [108,109,110,111,112]. This suggests the importance of model selection based on understanding the problem, rather than an automatic preference for the most complex model available.

Two cross-cutting problems tend to repeat: (i) data-related issues (limited annotations, heterogeneous sensors, domain shifts) that restrict generalization; and (ii) the problem of interpretability for truly trustworthy recommendations at the farm and policy levels. The latest explainable AI and model interpretation work points to both the opportunities and the dangers of black-box adoption for climate ramifications [51,55,113]. The resolution of these problems will demand diligent benchmarking and calibrating & validating frameworks, especially for spatial and temporal resolution that fundamentally influence error and bias [51,111,113].

Increasingly common are hybrid approaches, e., integration of process-based models and ML or the combination of optical, SAR, and meteorological streams—which steadily enhance the accuracy and transferability of models to different environments [37,40,61,63,73,100,114,115]. These frameworks also facilitate counterfactual reasoning (e.g., management changes) while preserving the pattern-mining capabilities of ML during cross-sectional analysis.

The AI literature reviewed shows that AI enhances the timeliness and precision of yield intelligence and improves production sustainability and resilience when supported by good data practice and good agronomic context. I recommend that future research focus on: (1) developing publicly available and curated comprehensive datasets that encompass different climates and agricultural management practices; (2) developing explainable and uncertainty-aware decision support models; and (3) the creation of hybrid, multi-sensor approaches that are designed for operational scalability within areas of limited data resources.