AgriST-Trans: A self-supervised Transformer pre-trained model for crop classification based on time-series remote sensing

Benefiting from the development of a new generation of earth observation satellites, the number of highquality remote sensing images has increased substantially. In agriculture, using remote sensing images for crop classification is an important application. However, despite the existence of a large amount of unannotated remote sensing data, routine supervised learning methods are limited by data annotation requirements, making it difficult to fully utilize these data. Therefore, the research and development of new methods to effectively use unannotated remote sensing data is of great significance to the agricultural field. In this paper, a novel self-supervised pre-training model AgriST-Trans, which employs a Transformer encoder as the base model and is pre-trained with a large number of accessible unlabeled multi-band Sentinel-2 time series remote sensing images. To fully utilize temporal information, the model takes time series of image blocks as input. The model extracts dynamic information on crop growth from remote sensing images of multiple periods, providing additional discriminative features for crop identification and integrating spatiotemporal information to improve the accuracy of crop classification. At the same time, according to the principle of mask autoencoder, the image block is randomly masked, so that the model can build the unknown image block through the known image block, so as to force the model to learn the rich spatial, spectral, and temporal characteristics in the remote sensing image. After completing the pre-training, the pre-training network can be further adjusted on the marker data of the related tasks to migrate the common features learned from the satellite image time series to the corresponding tasks. This approach helps to improve the generalization performance of the model, especially when label data is scarce. To verify the effect of the pre-trained model. In this study, three regions of the U.S. and 2021 crop data layers were selected as the crop categorization and labeling dataset. The experimental results demonstrate the validity of this study, the proposed AgriST-Trans model. Compared to the purely supervised learning model, the method of this study significantly improves the identification accuracy, demonstrating the potential of this method in the practical application of agricultural geographic information.