Development and validation of a crop yield prediction framework accounting for regional heterogeneity: A case study of spring maize in Jilin Province

Abstract

Accurate and timely crop-yield prediction is essential for ensuring food security, managing agricultural risk, and supporting policy formulation. To address the respective limitations of traditional crop growth models and deep learning methods under complex environmental conditions, a hybrid modeling framework is proposed that integrates remote-sensing data assimilation, a process-based crop growth model, and deep learning techniques. Using spring maize in Jilin Province, China (2015–2020) as the case study, leaf area index (LAI) retrieval accuracy is first improved by coupling the PROSAIL model with machine-learning algorithms. A complete meteorological sequence for the target year is then constructed using a dynamic time warping (DTW) algorithm to overcome early-season prediction challenges caused by missing real-time weather data. Retrieved LAI is assimilated into the WOFOST crop growth model through an ensemble Kalman filter (ENKF) to calibrate state variables and enable dynamic yield prediction across growth stages. Finally, a deep learning model (Convolutional Neural Network–Attention Long Short-Term Memory with Multi-Task Learning, CNN-ALSTM-MTL) is developed to fuse assimilation outputs with multi-source heterogeneous data, leveraging multi-task learning to enhance adaptability to regional heterogeneity and improve yield prediction performance at the regional scale. Assimilation is found to substantially improve maize-yield estimation, increasing R² by 0.2 and reducing RMSE by 276 kg ha⁻¹. Compared with the assimilated crop growth model alone, the hybrid framework further increases R² by 35 % and decreases RMSE by 23 % by hierarchically capturing feature information relevant to maize-yield estimation. The best performance is achieved during the key growth stage (jointing to tasseling stage), with an R² of 0.75 and an RMSE of 592 kg ha⁻¹, enabling reliable yield prediction approximately two months before harvest. This framework demonstrates potential for cross-crop and cross-regional applications and provides robust methodological support for regional-scale yield forecasting and food-security early warning.

Plant phenotyping relevance

LAIという植物形質の推定・同化と収量推定を中核とする統合的な計算フェノタイピング／予測フレームワークを開発・検証しており、単なる農業実験での routine 測定ではない。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.