Mid-Season Wheat Yield Estimation and Decision Support: Bridging Regularized Machine Learning and Full-Stack Geospatial Web Architecture

Source abstract

Background: Wheat (Triticum aestivum) is a cornerstone of national food security in Pakistan, par-ticularly in Punjab province. However, climate variability, rising temperatures, and unpredictable precipitation patterns increasingly threaten yields. Traditional crop-yield forecasting relies heavily on manual, post-harvest reporting that leaves government authorities and farming communities without operational visibility during the growing season. Objective: This study evaluates whether regularized regression can accurately forecast season-level wheat yields using multispectral satellite and climate data restricted to the first 50% of the growing season (mid-February cutoff), thereby providing a critical eight-week lead time before harvest. Furthermore, it presents the engineering and deployment of an operational full-stack web application that translates this predictive model into a real-time decision-support system. Methods: We analyzed a six-season dataset (2019–2020 to 2024–2025) across experimental fields at the University of Agriculture Faisalabad (UAF) PARS Campus. High-resolution Sentinel-2 Normalized Difference Vegetation Index (NDVI) observations, CHIRPS daily precipitation estimates, and ERA5-Land temperature reanalysis records were integrated to derive 25 agroclimatic and phenological features. Six regression algorithms—ElasticNet, Ridge, Partial Least Squares (PLS), Principal Component Regression (PCR), Support Vector Regression with a Radial Basis Function (SVR-RBF), and Gaussian Process Regression (GPR)—were evaluated under a strict Leave-One-Season-Out Cross-Validation (LOOCV) framework. Feature selection was conducted strictly in-fold to prevent data leakage. Results: The optimized ElasticNet model achieved the highest predictive stability and accuracy at the 50% season-elapsed window, yielding an out-of-fold R2 of 0.868, a Root Mean Squared Error (RMSE) of 64.72 kg, a Mean Absolute Error (MAE) of 59.14 kg, and a Mean Absolute Percentage Error (MAPE) of 3.36%. Comparative analyses showed that ElasticNet significantly outperformed high-complexity, non-regularized models, which suffered severe overfitting due to the high-dimensional, small-sample (n = 6) nature of season-level crop datasets. Growing De-gree Days (GDD) and GDD–NDVI interaction terms were identified as the primary predictors of final productivity. Conclusions: To operationalize these findings, a production-grade web application was developed utilizing a FastAPI (ASGI) backend and a React.js frontend. By integrating the Google Earth Engine (GEE) REST API, the platform enables real-time, cloud-based data extraction and prediction for custom user-drawn field polygons, with predictions stored in a MongoDB NoSQL database with GeoJSON spatial indexing. The application is actively deployed and utilized by agricultural extension officers, researchers, and regional policymakers.