Enhanced Crop Mapping Using Polarimetric SAR Features and Time Series Deep Learning: A Case Study in Bei'an, China

文献类型： 外文期刊

第一作者： Liu, Niantang

作者： Liu, Niantang;Duan, Si-Bo;Liu, Xiangyang;Zhao, Qunshan;Williams, Richard;Barrett, Brian

作者机构：

关键词： Crops; Feature extraction; Accuracy; Deep learning; Synthetic aperture radar; Remote sensing; Satellites; Monitoring; Long short term memory; Data mining; Crop mapping; deep learning; multitemporal satellite data; polarimetric synthetic aperture radar (PolSAR)

期刊名称：IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING （ 影响因子：8.6； 五年影响因子：8.8 ）

ISSN： 0196-2892

年卷期： 2025 年 63 卷

页码：

收录情况： SCI

摘要： Large-scale crop mapping is essential for decision-makers to evaluate agricultural resource usage and estimate crop yields. Considering the utility of annual crop inventory (CI) statistics for monitoring crop growth, generalizing near real-time crop classification over large areas becomes necessary. Accurate crop-type identification using remote sensing data remains challenging due to the variability in crop growth patterns across time and space, the presence of crops with similar phenological stages, and the scarcity of labeled data. This study develops deep learning-based approaches to map agricultural regions at the county level using multitemporal Sentinel-1 synthetic aperture radar (SAR) data, specifically evaluating the contribution of SAR-derived input predictors for discriminating both majority and minority crops in Bei'an, Northeast China. The proposed model architecture amalgamates 1-D convolutional layers (Conv1D) with attention-based long short-term memory (LSTM) to characterize the crop types exhibiting phenological similarities using a range of SAR-derived input predictors. The results are compared with alternative multitemporal deep learning frameworks, including standalone Conv1D and Transformer models, as well as the machine learning algorithm random forest (RF), which serves as the baseline for comparison. The designed architecture (Conv1D-LSTM) achieved the highest $F1$ scores (maize: 87%, soybean: 86%, and other crops: 85%) when applied to an inherently imbalanced dataset, using m-chi decomposition features as input predictors. The results provide superior performance in terms of effectiveness and efficiency compared to other selected models. The monthly in-season crop classification underscores the importance of temporal dependencies and the availability of multitemporal observations for learning dynamic growth patterns over large areas. Furthermore, the interpretation of model learning processes and outcomes is explained by visualizing weight distributions and hidden features. This study offers a comprehensive evaluation of essential SAR features in multitemporal satellite data for accurate crop mapping, utilizing advanced deep learning techniques. This work is available at https://github.com/Niantangliu/Deep-learning-crop-mapping.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序