基于改进DenseNet模型的高分遥感影像城市裸地提取

利用遥感影像提取裸地是监测裸地空间分布的一个重要手段。针对目前普遍存在的边界不清晰、空间信息丢失、小面积裸地漏提和与高反射率建筑不易区分等问题，设计了一种改进DenseNet的遥感裸地提取深度学习模型，主要采取密集连接块、坐标卷积和密集空洞空间金字塔3种方法，增强DenseNet模型在获取坐标信息、丰富裸地空间特征信息、对全局上下文信息感知等方面的能力，减少模型对于空间细节特征丢失环节，提高裸地遥感提取的精度。实验表明，该方法提取裸地的总精度为97.66％、交并比为68.69％、综合评价指标F1为81.44％、召回率为76.62%以及虚警率为25.68%，明显优于其他机器学习方法和深度学习方法。此外，该模型对于多源遥感影像上的裸地提取也具有良好的普适性，在高分一号、高分六号和哨兵二号等遥感数据集上测试的总精度分别为95.80％、93.00％和92.55％；交并比分别为75.18％、75.13％和50.47％；综合评价指标分别为85.83％、85.80％和67.08％。因此，改进的DenseNet模型方法较其他方法更适用于裸地的提取。

Extraction of urban bare land from high-resolution remote sensing images based on improved DenseNet model

The extraction of bare land using remote sensing images is an important means to monitor the spatial distribution of bare land. Aiming at the problems of unclear boundaries, spatial information loss, small area bare land missing extraction and difficulty in distinguishing buildings with high reflectivity, a deep learning model from remote sensing bare land extraction with improved DenseNet was designed, mainly adopting three methods of DenseBlock, coordinate convolution and dense atrous spatial pyramid pooling to enhance the ability of DenseNet model in acquiring coordinate information, enriching spatial feature information of bare land and sensing global context information, reducing the model for spatial detail feature loss links and improving the accuracy of bare land remote sensing extraction. The experiments showed that the method extracts bare land with 97.66% overall accuracy, the IoU was 68.69%, the comprehensive evaluation index F1 was 81.44%, and the recall rate was 76.62. %, the false alarm rate was 25.68%, which was significantly better than other machine learning methods and deep learning methods. In addition, the model also has good universality for bare ground extraction on multi-source remote sensing images. The OA tested on remote sensing datasets such as GF-1, GF-6 and Sentinel-2 were 95.80%, 93.00%, and 92.55%, respectively; the IoU were 75.18%, 75.13%, and 50.47%, respectively; the F1 were 85.83%, 85.80% and 67.08%, respectively. Therefore, the improved DenseNet model method is more suitable for the extraction of bare land than other methods.