Estimating the full-period rice leaf area index using CNN-LSTM-Attention and multispectral images from unmanned...

Leaf area index (LAI) of rice is a crucial parameter for assessing the growth conditions and predicting yields. However, traditional measurement methods are inefficient and insufficient for large-scale monitoring.

This study proposes a CNN-LSTM-Attention (CLA) model that integrates convolutional neural networks (CNN), long short-term memory (LSTM), and a self-attention mechanism, aiming to achieve high-precision estimation of rice LAI across all growth stages based on the unmanned aerial vehicle (UAV) multispectral imagery and deep learning techniques. The estimation performance of vegetation indices (VIs), machine learning methods (SVR, RFR, PLSR, XGBoost), and deep learning models (DNN, CNN, LSTM) were comparatively analyzed.

The results show that the CLA model outperforms other approaches in estimating rice LAI throughout the entire growing period, achieving a coefficient of determination (R²) of 0.92 and a relative root mean square error (RRMSE) below 9%, significantly better than linear regression and machine learning methods. Moreover, the CLA model maintains high stability and accuracy across different LAI ranges, with notably reduced errors for low LAI values (one to three), effectively mitigating the influence of soil background. This research offers an efficient and accurate technological approach for rice growth monitoring and holds significant implications for precision agricultural management.