Large Scale and High-Precision Crop Recognition Using Satellite Remote Sensing Images

Precise large-scale crop recognition is crucial for estimating crop yield and carbon assimilation. Current research on crop type recognition based on deep learning is often limited to a single crop and small areas, with poor model transferability. When performing multi-crop recognition over large areas, the accuracy significantly decreases. Therefore, it is necessary to establish a high-precision crop type recognition model with strong generalization capability. A deep learning model (LSTM-Transformer Parallel, LTP) based on the combination of LSTM (Long Short-Term Memory) network and Transformer self-attention mechanism architecture is proposed in this paper. The model captures the dependencies among crop NDVI time series data to achieve high-precision recognition of large-scale crop types. Experimental results show that in crop recognition experiments using small-scale agricultural area data, the recognition results of the LTP, TSMixer, and SegRNN models all performed excellently, with average accuracies of 94.55%, 94.41%, and 94.79%, respectively. In contrast, in crop recognition experiments using large-scale agricultural area data, the average accuracy of the LTP model reaches 93.77%, while the accuracies of the TSMixer and SegRNN models are 82.30% and 84.13%, respectively. This indicates that the LTP model not only performs well on small-scale agricultural area datasets but also significantly outperforms models in the literature on large-scale agricultural area datasets, demonstrating strong generalization capability. It provides a reliable method for large-scale multi-crop recognition, contributing to global food security and carbon cycle assessment.