Scholarlink

<p>Satellite imagery plays a crucial role in crop monitoring. While the U-Net network excels in optical remote sensing image analysis, it struggles with complex high-dimensional microwave remote sensing data, such as Synthetic Aperture Radar (SAR). SAR provides high-precision geospatial data as an all-weather observation method, but existing U-Net frameworks fail to effectively capture complex spatiotemporal features in high-resolution time-series imagery, compromising crop planting distribution extraction accuracy. To address this, we propose a Spatiotemporal Intelligent Augmentation and Extension Network (ST-IAE-Net), which enhances performance by optimizing the U-Net architecture and loss mechanism. The framework adjusts output channels to better represent SAR image details and incorporates a multi-scale convolutional structure to improve temporal feature extraction. By integrating a Spatiotemporal Self-Attention (STSA) module and a Multi-scale Cross-Attention (MCA) mechanism, it effectively aggregates multi-scale spatiotemporal information. Additionally, we innovatively propose a hybrid loss function combining Dice Loss (to handle class imbalance) and Focal Loss (to focus on hard-to-classify samples), with weight optimization performed using Optuna. Experiments conducted on datasets from the Arkansas River Basin (USA) and the Weihe Plain (China) demonstrate that ST-IAE-Net improves overall accuracy (OA), precision, recall, and F1-score by 5.1, 6.3, 4.8, and 5.7%, respectively, compared to baseline models. Furthermore, the hybrid loss function enhances these metrics by an additional 2.2, 2.1, 2.2, and 2.2%, significantly boosting the accuracy of crop planting distribution extraction.</p> Graphical abstract <p>In this study, we selected SAR data from the Arkansas River Basin in the United States, spanning from April to June between 2018 and 2023, along with corresponding CDL ground truth labels. After preprocessing, single-temporal data from each month (April to June) were synthesized annually. Utilizing this multi-temporal dataset, Sentinel-1 synthetic mosaic images were segmented into 256 256 pixel patches. These patches were divided such that 80% of the annual image collection served as training data, 5% as validation data, and 15% as testing data. The ST-IAE-Net network was trained on the training dataset, and upon completion, the model’s performance and hyperparameter tuning were evaluated using the validation dataset. Subsequently, the SAR data from the test set were input into the model to obtain predicted winter wheat classification results. Following this, SAR data from April to June 2023 in the Weihe Plain of Shaanxi Province, China, were preprocessed and input into the model. The model’s classification outputs were then compared with the corresponding ground truth labels from Shaanxi Province, China, to assess the model’s accuracy. This process further validated the model’s transferability and robustness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A New Approach for Wheat Planting Area Extraction with Remote Sensing Data

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Yunlong Du,
  • Jiaqian Lian,
  • Xinyue Duan,
  • Xiaofei Kuang,
  • Pengliang Wei,
  • Jiao Guo

摘要

Satellite imagery plays a crucial role in crop monitoring. While the U-Net network excels in optical remote sensing image analysis, it struggles with complex high-dimensional microwave remote sensing data, such as Synthetic Aperture Radar (SAR). SAR provides high-precision geospatial data as an all-weather observation method, but existing U-Net frameworks fail to effectively capture complex spatiotemporal features in high-resolution time-series imagery, compromising crop planting distribution extraction accuracy. To address this, we propose a Spatiotemporal Intelligent Augmentation and Extension Network (ST-IAE-Net), which enhances performance by optimizing the U-Net architecture and loss mechanism. The framework adjusts output channels to better represent SAR image details and incorporates a multi-scale convolutional structure to improve temporal feature extraction. By integrating a Spatiotemporal Self-Attention (STSA) module and a Multi-scale Cross-Attention (MCA) mechanism, it effectively aggregates multi-scale spatiotemporal information. Additionally, we innovatively propose a hybrid loss function combining Dice Loss (to handle class imbalance) and Focal Loss (to focus on hard-to-classify samples), with weight optimization performed using Optuna. Experiments conducted on datasets from the Arkansas River Basin (USA) and the Weihe Plain (China) demonstrate that ST-IAE-Net improves overall accuracy (OA), precision, recall, and F1-score by 5.1, 6.3, 4.8, and 5.7%, respectively, compared to baseline models. Furthermore, the hybrid loss function enhances these metrics by an additional 2.2, 2.1, 2.2, and 2.2%, significantly boosting the accuracy of crop planting distribution extraction.

Graphical abstract

In this study, we selected SAR data from the Arkansas River Basin in the United States, spanning from April to June between 2018 and 2023, along with corresponding CDL ground truth labels. After preprocessing, single-temporal data from each month (April to June) were synthesized annually. Utilizing this multi-temporal dataset, Sentinel-1 synthetic mosaic images were segmented into 256 256 pixel patches. These patches were divided such that 80% of the annual image collection served as training data, 5% as validation data, and 15% as testing data. The ST-IAE-Net network was trained on the training dataset, and upon completion, the model’s performance and hyperparameter tuning were evaluated using the validation dataset. Subsequently, the SAR data from the test set were input into the model to obtain predicted winter wheat classification results. Following this, SAR data from April to June 2023 in the Weihe Plain of Shaanxi Province, China, were preprocessed and input into the model. The model’s classification outputs were then compared with the corresponding ground truth labels from Shaanxi Province, China, to assess the model’s accuracy. This process further validated the model’s transferability and robustness.