SMN-AgroCLA: Comparison of Different Normalization Methods for Improving Rice Yield Prediction Accuracy Using Remote Sensing Data in Eastern China from 2008 to 2017

Le résumé fourni par la source

Abstract. Yield prediction is crucial for national food security and the formulation of trade policies. Most deep learning (DL) models rely on normalization methods to process input data, aiming to enhance the stability of model training and accelerate convergence speed. However, the importance of data preprocessing (i.e., input data normalization) in DL-based yield prediction is underemphasized. Furthermore, conventional methods fail to address distortions in feature scaling caused by extreme values, such as abnormally high precipitation, leading to increased prediction errors. In this study, we proposed a Sequential Midrange Normalization (SMN) method and integrated it with the newly designed Agricultural-CNN-LSTM-Attention (AgroCLA) model, collectively termed the SMN-AgroCLA framework, to improve rice yield prediction accuracy under extreme weather conditions. To validate the efficacy of the SMN, we compared it with four other commonly used normalization methods and conducted yield prediction experiments across six different DL models, by using Moderate Resolution Imaging Spectroradiometer, Global Precipitation Measurement and other multi-source remote sensing data of the Eastern China from 2008–2017. The results shown that SMN method consistently outperformed superior yield prediction performance even in years affected by extreme meteorological disasters (e.g., 2015), achieving an R² of 0.815, which was 17.3 % higher than the next best method, ZSN (Z-Score Normalization). Based on SMN, the accuracy and generalization of all models were optimized, with the AgroCLA achieved the highest accuracy (with R²=0.841). Additionally, the model's performance peaked around the flowering stage (around mid-August, R² =0.859), which is two months ahead of the harvest season. This study demonstrates the critical role of data normalization in deep learning-based yield prediction and offers a practical solution to mitigate the threat of increasing extreme meteorological disasters to food security.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.