USTAnet: A Neural Network for Enhanced Resolution Generalization in Kilometer-Level Rice Yield Prediction

Rice, as the main food crop for half of the world’s population, could directly threaten food supply chain stability if precise yield prediction methods are lacking. However, existing deep learning (DL) models exhibit strong scale dependency: when the input data resolution changes, the scale of spatial features extracted by the model no longer aligns with the actual size of target objects, leading to performance degradation due to mismatched receptive fields. To address the challenge posed by spatial resolution variation on the generalization ability of remote sensing DL models, this study proposes a DL model, U-shaped spatial–temporal attention network (USTAnet), by designing the multiscale convolution (MSC) block and multidimensional feature integrator blocks. These designs contribute to USTAnet’s enhanced resolution generalization (the model’s adaptability to remote sensing data with different spatial resolutions) and spatiotemporal generalization capabilities. USTAnet outperforms existing models in accuracy and generalization ability, including U-Net, CNN-GRU, LSTM-CNN, SSTNN, and DDCN, with an R$^{2}$ of 0.817 at a 4 km × 4 km resolution. In 1 km × 1 km resolution generalization tests, USTAnet consistently shows the best performance across all years (2011–2015), with an average R$^{2}$ of 0.758, which is 15% higher on average than other models. In spatial–temporal generalization tests, USTAnet exhibits the highest performance in 2015, with R$^{2}$ values of 0.817. This study proved that USTAnet is a promising kilometer-level yield prediction model with great potential for cross-resolution rice yield prediction applications, which can help promote the application of remote sensing technology in agricultural production.