Chang Chen, Chao Zhang, Lihua Zhao, Cuicui Yang, Xiaochuang Yao, Bin Fu
Abstract
Accurate crop spatial distribution data are a fundamental basis for optimizing cropping structures and enhancing grain production. Timely and precise crop classification mapping is essential for effective agricultural monitoring. This study proposes a multi-module remote sensing classification model, the GEE-Deep-Crop-Mapping model (GDCM), constructed on the basis of multi-source remote sensing data from the Google Earth Engine (GEE) platform. The GDCM consists of two main modules: a multi-scale spatial module integrating dilated and depthwise separable convolutions with CBAM to enhance spectral-textural features across varying field sizes; A temporal dynamics module combining Bi-LSTM and multi-head attention to capture phenological dependencies and identify critical growth phases. The GDCM is Applied to Nenjiang City, Heilongjiang province, China with Sentinel-1 and Sentinel-2 time-series data, the GDCM model achieved over 95% accuracy, outperforming both the Bi-LSTM and Random Forest (RF) models. Similarly, the model achieved 90% accuracy in Cass County, USA. Additionally, by leveraging transfer learning, the optimized model successfully extracted multi-year crop distribution data, facilitating the monitoring and analysis of grain-soybean rotation patterns from 2017 to 2022 in Nenjiang City. These findings highlight the potential of the GDCM model for advancing agricultural monitoring and optimizing cropping structures.
1-s2.0-S0168169925010658-main.pdf