A method of rice yield prediction based on the QRBILSTM-MHSA network and hyperspectral image

Abstract

Accurate and timely prediction of rice yield is crucial for ensuring food security and optimizing agricultural management. This study proposes a novel QRBILSTM-MHSA model (Quantile Regression-based Bidirectional Long Short-Term Memory Network with Multi-Head Self-Attention) for rice yield prediction, synergizing hyperspectral imaging with multi-modal phenotypic data. The model replaces traditional RNN architectures with BILSTM to acquire bidirectional temporal patterns and permanent dependencies in rice growth cycle. A multi-head self-attention (MHSA) is introduced to weight critical growth factors through parallel subspace analysis, while quantile regression (QR) provides interval predictions, simultaneously estimating average yield and fluctuation ranges. Experimental results demonstrate that the proposed model achieves an R2 of 0.927, a MAPE of 2.21%, and an RMSE of 0.22 tons/ha, significantly outperforming traditional methods such as LSTM, BP-NN, RF, SVR, and ARIMA. At a 95% confidence level, the model achieves a prediction interval coverage probability (PICP) of 98.8% and a percentage of interval width mean percentage (PIWMP) of 0.16, indicating high reliability and robustness. This study highlights the potential of integrating hyperspectral data and deep learning for precise and scalable rice yield prediction, offering valuable insights for agricultural decision-making.

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.