关键词: GF-1, 多光谱, 碍洪物, 人工智能, 识别模型

Abstract: The obstructions in waterways are significant factors affecting flood disasters, thus their efficient and precise management has garnered widespread attention. Traditional manual inspections cannot meet the needs for efficient and precise applications; the integration of artificial intelligence (AI) with remote sensing technology applications is an inevitable path. However, the performance of many AI models in remote sensing applications is not yet clear and urgently needs further investigation. This research takes the Datengxia reservoir area in Guangxi as an example to study the construction methods of AI recognition models for obstructions in waterways using remote sensing. Based on GF-1 remote sensing imagery, an obstructions in waterways training dataset is constructed. Using ResNet101 as the core Network, six current mainstream semantic segmentation models are adopted, including PSPNet, PAN, MANet, FPN, DeepLabV3+, and UNet++. The models are trained for the identification of obstructions in waterways to further evaluate their precision and efficiency. Key findings include: ①Deep learning models utilize ResNet101 as the backbone Network shows excellent performance in identifying obstructions in waterways with all models achieving an F1 score above 0.70 and IoU above 0.58. Among them, the DeepLabV3+ model, which combines atrous convolution and global pooling techniques, achieves an F1 score of 0.82 and an IoU of 0.72, demonstrating significant advantages in capturing contextual information and micro-features. ②PSPNet, despite having a lower number of parameters, exhibits high processing efficiency and accuracy, capable of handling 8 samples per batch with a frame rate of 10.49. In summary, DeepLabV3+ stands out in precise identification and contour delineation, while PSPNet shows great potential in large-scale data processing. The study results can provide a reference for constructing AI remote sensing models and offer technical support for waterway safety monitoring.

Key words: GF-1, multispectral, flood obstruction objects, artificial intelligence, recognition model