SMN-AgroCLA: a deep learning framework with sequential midrange normalization for enhanced rice yield prediction using remote sensing data

Résumé fourni par la source

Yield prediction is crucial for ensuring national food security and informing trade policies. Most deep learning (DL) models employ normalization techniques to preprocess input data, with the goal of improving training stability and accelerating convergence. However, the role of data preprocessing (i.e. input data normalization) in DL-based yield prediction remains underemphasized. Moreover, conventional normalization approaches often struggle to handle distortions in feature scaling caused by extreme values, such as unusually high precipitation, which can lead to increased prediction inaccuracies. In this study, we introduce a Sequential Midrange Normalization (SMN) method and combine it with the newly developed Agricultural-CNN-LSTM-Attention (AgroCLA) model. This integrated framework, referred to as SMN-AgroCLA, is designed to enhance the accuracy of rice yield predictions under extreme weather conditions. To validate the efficacy of the proposed SMN, we compared it against four other widely used normalization techniques. Yield prediction experiments were conducted using six different deep learning models, incorporating multi-source remote sensing data – including Moderate Resolution Imaging Spectroradiometer (MODIS) and Global Precipitation Measurement (GPM) – from Eastern China between 2008 and 2017. The results demonstrated that SMN method consistently delivered superior prediction performance, even in extreme meteorological conditions such as those experienced in 2015. It achieved an R² of 0.815, representing a 17.3% improvement over the next best method, Z-Score Normalization (ZSN). Furthermore, when integrated with SMN, all models exhibited enhanced accuracy and generalization capability, with the AgroCLA achieving the highest accuracy (with R² = 0.841). Model accuracy peaked around the flowering stage (around mid-August, R² = 0.859), approximately two months ahead of harvest. 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.