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.

中文翻译：

---

用于时间序列遥感影像中分散小农田作物类型制图的双分支网络


随着遥感技术的飞速发展，利用时间序列遥感影像识别农田地块已成为一项日益重要的任务。在本文中，我们专注于识别许多亚洲地区分散、不规则和定义不明确的农田中的作物。我们选择了两个具有小地块和散散地块的代表性位置，并构建了两个新的时间序列遥感数据集 （JM 数据集和 CF 数据集）。我们提出了一种新的深度学习模型 DBL，即具有长短期记忆 （LSTM） 的双分支模型，它利用主分支和补充分支来完成准确的作物类型映射。主分支设计用于捕获全局感受野，而 supplementary 分支用于时间和空间特征细化。进行实验是为了评估 DBL 与最先进的 （SOTA） 模型相比的性能。结果表明，DBL 模型在这两个数据集上都表现得非常出色。特别是在以分散和不规则图为特征的 CF 数据集上，DBL 模型实现了 97.70% 的总体准确率 （OA） 和 90.70% 的平均交并比 （mIoU）。它的性能优于所有 SOTA 模型，并成为唯一一个 mIoU 分数超过 90% 的模型。我们还展示了 DBL 在不同农业地区的稳定性和稳健性。