A dual-branch network for crop-type mapping of scattered small agricultural fields in time series remote sensing images

With the rapid advancement of remote sensing technology, the recognition of agricultural field parcels using time-series remote sensing images has become an increasingly emphasized task. In this paper, we focus on identifying crops within scattered, irregular, and poorly defined agricultural fields in many Asian regions. We select two representative locations with small and scattered parcels and construct two new time-series remote sensing datasets (JM dataset and CF dataset). We propose a novel deep learning model DBL, the Dual-Branch Model with Long Short-Term Memory (LSTM), which utilizes main branch and supplementary branch to accomplish accurate crop-type mapping. The main branch is designed for capturing global receptive field and the supplementary is designed for temporal and spatial feature refinement. The experiments are conducted to evaluate the performance of the DBL compared with the state-of-the-art (SOTA) models. The results indicate that the DBL model performs exceptionally well on both datasets. Especially on the CF dataset characterized by scattered and irregular plots, the DBL model achieves an overall accuracy (OA) of 97.70% and a mean intersection over union (mIoU) of 90.70%. It outperforms all the SOTA models and becomes the only model to exceed 90% mark on the mIoU score. We also demonstrate the stability and robustness of the DBL across different agricultural regions.