Crop spectral data reconstruction and classification based on temporal masked autoencoder

Satellite imagery holds immense potential for crop monitoring due to its wide coverage and long-term stable historical data. However, frequent cloudy and rainy weather results in extremely fragmented optical remote sensing data in the temporal dimension, creating numerous observation gaps. This study proposes a temporal masked auto-encoder (T-MAE) framework that treats cloud occlusion as a natural mask. By performing self-supervised pre-training on large-scale unlabeled Sentinel-2 imagery, the model is forced to learn the intrinsic temporal dependencies and spectral evolution patterns of crop growth. Furthermore, the reliability of the reconstructed spectral data is evaluated by generating plot-level crop type maps using the reconstructed spectral time-series. The research results indicate that: T-MAE can reconstruct complete crop growth curves with high precision even under extreme conditions with only 20% valid observations. In downstream classification tasks, the classifier based on T-MAE pre-trained features achieved higher accuracy compared to bidirectional long short-term memory (Bi-LSTM) and Temporal Convolutional Neural Network (Temp-CNN) models (which rely on linear interpolation and Whittaker smoothing algorithms to handle cloud occlusion). Moreover, the model pre-trained with a 75% masking rate yielded higher classification accuracy than those pre-trained with 25% and 90% masking rates. In conclusion, T-MAE not only outperforms existing methods in crop classification accuracy but also demonstrates superior spatiotemporal generalization and robustness against interference. This work provides a new paradigm for addressing the dual challenges of label scarcity and cloud interference in agricultural remote sensing.