Light is shed on the effect of dissimilarity between train and test feature distributions on forecasting models, compares deep learning versus non-deep learning models, and introduces modifications that are effective for non-deep learning models.
Abstract
The impact of climate variability on food production has led to the creation of various forecasting models that uses machine learning (ML), numerical weather predictors (NWP) or a hybrid of ML-NWP models to identify structural and physical relationships between meteorological drivers and crop growth, in order to predict crop yield. Droughts, for example the 2012 Midwestern US (Corn Belt) drought, are extreme events that affect crop production and test the limits of these forecasting models. Using 16 meteorological drivers as predictors, we compare ML (non-deep learning) and deep learning forecasting models to predict the county-level corn yield for the extreme drought year, 2012. This forecasting problem is characterized by a dissimilarity between the feature distributions of the training and test data, where the meteorological conditions of the extreme drought year fall outside the range of historically observed values. Additionally, the dataset consists of spatial and temporal irregularities where counties with missing yields introduce spatial sparsity and the use of only a subset of daily values per year introduce temporal sparsity. To overcome this, we use sample weighting and feature selection as modifications to improve our forecasting models. These modifications lead to an improvement for ML models; however, the deep learning model VITA shows little to no improvement. While VITA outperforms the ML models with or without modifications, our current study sheds light on the effect of dissimilarity between train and test feature distributions on forecasting models, compares deep learning versus non-deep learning models, and introduces modifications that are effective for non-deep learning models.
The increasing availability of agricultural timeseries data enabled more accurate and data-driven crop yield prediction. However, raw meteorological, soil, and vegetation datasets often fail to capture complex temporal dependencies essential for robust forecasting. This study proposes a structured feature engineering framework for crop yield prediction using multi-source agricultural time-series data, including climatic variables, soil properties, and satellite-derived vegetation indices. Temporal features such as lag variables, rolling statistics, cumulative rainfall indices, and growing degree days (GDD) are systematically extracted to enhance model interpretability and predictive performance. Multiple machine learning (ML) models, including random forest (RF), gradient boosting (GH), support vector regression (SVR), and long short-term memory (LSTM) networks, are evaluated. Experimental results demonstrate that engineered temporal features reduce MAE and RMSE by approximately 25–35% compared to raw feature baselines. The optimized LSTM model achieved an RMSE of 3.96 tons/ha and r
2
of 0.92, outperforming traditional regression models. Feature importance analysis confirms the significant contribution of cumulative rainfall, temperature lags, NDVI (Normalized Difference Vegetation Index) trends, and soil moisture dynamics. The proposed framework provides a scalable and adaptable solution for precision agriculture, enabling improved yield forecasting and data-driven decision-making under climate variability.
K. Lata, A. Kumar, P. Sharma et al.· BIO Web of Conferences· 0 citations
This study presents a geospatially informed machine learning approach to improve crop yield prediction in Zambia, where agriculture underpins rural livelihoods and national food security. The research integrates satellite-derived vegetation indices, principally the Normalised Difference Vegetation Index (NDVI), with meteorological indicators including seasonal rainfall distribution and temperature trends, across a 25-year wheat yield record (1999 to 2024) for a commercial farm in Chongwe District, Zambia. Datasets were harmonised through spatial standardisation, feature engineering, and temporal aggregation to produce a coherent input structure for a Random Forest regression model benchmarked against Extreme Gradient Boosting (XGBoost). Rainfall frequency, seasonal thermal accumulation, and vegetation vigour emerged as the most influential predictors of yield variability. Random Forest achieved a stable coefficient of determination (R2 = 0.389), while XGBoost exhibited apparent superiority (R2 = 0.950) attributable to overfitting on a small, spatially homogeneous dataset. The findings demonstrate that model selection is critical in small-sample agricultural prediction contexts and contribute an interactive, field-ready decision-support dashboard for precision agriculture in Zambia.
Izukanji Muuka, Kasongo Changwe, Musoka Nyongolo et al.· Proceedings of the internati...· 0 citations
Accurate crop-yield forecasting is difficult when only small official-statistics datasets are available. This study evaluates a reproducible pre-season regional forecasting workflow for wheat, maize, and sunflower in Poltava region, Ukraine, using harmonised AgroStats records for 2010–2024. The analysis is based on one annual regional time series per crop; it is therefore a benchmark of temporal forecasting from official statistics rather than a gridded yield-prediction or spatial-mapping study. Forecast accuracy is assessed with mean absolute error (MAE), root mean squared error, and mean absolute percentage error. To avoid information leakage, feature transformations, imputations, scaling, and hyperparameter choices are performed only within the corresponding historical training window under a forward temporal design: training set 2010–2018, validation set 2019–2021, and held-out test set 2022–2024. In addition to ElasticNet, XGBoost, and LightGBM, the study compares transparent baseline forecasting methods: Naive, FORECAST.LINEAR, LINEST, and autoregressive integrated moving averag. Under the conservative lag-only scenario, the best 2022–2024 test results are obtained for maize (LightGBM, MAE 0.69 t ha−1) and sunflower (LightGBM, MAE 0.04 t ha−1), whereas for wheat the linear-trend baseline remains slightly better (MAE 0.49 t ha−1 versus 0.54 t ha−1 for ElasticNet). Supplementary analyses show that extended lag structures can improve selected crops and that seasonal NASA Prediction of Worldwide Energy Resources climate aggregates improve maize forecasts (MAE 0.52 t ha−1) but not wheat or sunflower. SHapley Additive exPlanations are used descriptively to examine whether the selected models rely on agronomically plausible predictors. The findings should be interpreted as crop-specific evidence under a small annual dataset: machine learning does not guarantee superiority over simple baselines, but it can provide a reproducible comparison framework and useful gains for selected regional forecasting tasks.
O. Kopishynska, Mark Fedorchenko, Y. Utkin et al.· Environmental Research Commu...· 0 citations
ABSTRACT Accurate sugarcane yield forecasting is essential for food security, sustainable agriculture, and resource management. However, conventional methods often struggle to capture the complex interactions that influence sugarcane growth, leading to unreliable predictions. This study explores whether a hybrid approach, combining remote sensing data, process-based crop modeling (APSIM), and machine learning, can enhance yield predictions. The developed model combines APSIM-simulated variables, weather data, and vegetation indices to forecast end-season sugarcane yield in São Paulo, Brazil, for the period 2010-2020. Sixteen regression models were evaluated at the municipal level. Incorporating APSIM-simulated variables as features in machine-learning models reduced the root mean square error (RMSE) of predictions between 7.7 and 26.9%. The exclusion of certain features elucidated that vegetation indices had the least impact on yield predictions. Weather data alone exerted a more significant impact on forecasts when used in process-based models compared to being directly input into machine-learning algorithms. Thus, our hybrid approach outperformed traditional methods, offering more accurate predictions. These findings could significantly enhance agricultural practices, empower farmers with improved predictive tools, and contribute to global food security and sustainable agriculture.
D.A.V. Grubert, B. Orsi, I. Z. Gonçalves et al.· Revista Brasileira de Engenh...· 0 citations
Accurate crop yield prediction requires both precision and interpretability for practical agricultural decision-making. This study is the first to systematically benchmark linear regression (LR) against gradient boosting baselines (XGBoost, LightGBM) alongside random forest (RF), support vector machine (SVM), and artificial neural network (ANN) for Indian agricultural yield forecasting, demonstrating that interpretable models can match or outperform black-box approaches. Using a comprehensive Indian agricultural dataset spanning multiple crops, states, and growing seasons (2000–2022), comprising 58,000 records across 22 states and 35 crop varieties, we analyzed features including cultivated area, rainfall, fertilizer applications, and pesticide usage. LR achieved R2 of 0.401 0.02 across 10-fold cross-validation, MSE of 480,239, MAE of 139.50, and RMSE of 692.99, outperforming complex models while maintaining full transparency. Transparent models allow policymakers to understand the impact of rainfall and fertilizer use, enabling evidence-based resource allocation. Results demonstrate that model complexity does not guarantee superior agricultural predictions, and LR provides interpretable coefficients—such as embedding LR coefficients into advisory tools to guide fertilizer recommendations and irrigation scheduling-enabling actionable agronomic insights for sustainable farming practices.
Imran Ahmad, R. Khokale, Nisha Gongal· Bulletin of Electrical Engin...· 0 citations
Northwestern Nigeria, a semi-arid Sahelian environment, is highly vulnerable to rainfall variability due to its strong dependence on rain-fed agriculture. This study evaluated the predictive capability of Artificial Neural Network (ANN) and Random Forest (RF) machine learning models in forecasting annual rainfall variability and aggregate cereal crop yields using some large-scale climate oscillation indices, specifically the El Niño-Southern Oscillation (ENSO), North Atlantic Oscillation (NAO), and the Atlantic Multi-decadal Oscillation (AMO). Historical data (2000-2024) were sourced from the National Aeronautics and Space Administration (NASA) Prediction of Worldwide Energy Resource (POWER) database for rainfall, the National Agricultural Extension and Research Liaison Services (NAERLS) for crop yields, and the National Oceanic and Atmospheric Administration (NOAA) for climate indices. Data preprocessing included temporal aggregation, normalisation, and quality control procedures before model development. The models were trained and validated using a chronological 80:20 data split, while predictive performance was evaluated using the coefficient of determination (R2), Mean Squared Error (MSE), and Mean Absolute Error (MAE). Results showed that the ANN model achieved slightly superior rainfall prediction performance (R² = 0.811; MAE = 101.82 mm), while the RF model produced higher predictive accuracy for aggregate cereal crop yields (R² = 0.893; MAE = 0.04375 T/Ha). The findings demonstrate the effectiveness of machine learning approaches in modelling complex climate-crop interactions across northwestern Nigeria. However, limitations associated with the relatively short temporal dataset and the exclusion of agronomic variables are acknowledged. The study highlights the potential application of machine learning models in climate-informed agricultural forecasting, early warning systems, and climate adaptation planning within vulnerable semi-arid regions.
I. A. Tanko, T. Yahaya, Aishetu Abdulkadir et al.· Kaduna Journal of Geography· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
