Dams constructed by humans are important facilities for irrigation, flood control, and power generation. Recognizing the location and number of dams is crucial for studying the impact of human activities on ecosystem change. Although many countries and organizations have established their own dam datasets, it is only the tip of the iceberg of real dam construction. Therefore, effectively and accurately obtaining the geographic location of dams is still a significant problem to be solved. This article proposes an improved convolutional neural network (CNN) based framework to recognize global dams from high-resolution remotely sensed images. First, a dataset named the global dam detection dataset is built based on Google earth high-resolution images, and the dataset is used as the training and testing dataset for the CNN model. Second, an improved dam recognition method (HRLibra-RCNN) is proposed to detect dams on a global scale. Third, an application strategy for global dam recognition from large remote sensing images is established to recognize dams in seven regions around the world. Compared with two two-stage object recognition models (Faster-RCNN and Cascade-RCNN) and a single-stage target detection model (RetinaNet), the proposed method achieved the highest average precision of 79.4%, with the HRNet-40w backbone network structures achieving the highest average precision of 80.7%. The average precision of 70.8% and recall of 90.4% are achieved during the application stage. The dataset and framework developed in this study are the first attempts to combine remote sensing big data and the deep learning method to recognize dams at a global scale.

中文翻译：

---

使用卷积神经网络从高分辨率遥感图像识别全球大坝


人类修建的水坝是灌溉、防洪、发电的重要设施。认识水坝的位置和数量对于研究人类活动对生态系统变化的影响至关重要。尽管许多国家和组织建立了自己的大坝数据集，但这只是实际大坝建设的冰山一角。因此，有效、准确地获取大坝的地理位置仍然是一个亟待解决的重大问题。本文提出了一种基于改进的卷积神经网络（CNN）的框架，用于从高分辨率遥感图像中识别全球大坝。首先，基于Google Earth高分辨率图像构建了名为全球大坝检测数据集的数据集，并将该数据集作为CNN模型的训练和测试数据集。其次，提出了一种改进的水坝识别方法（HRLibra-RCNN）来检测全球范围内的水坝。第三，建立了大遥感图像全球大坝识别的应用策略，识别全球七个地区的大坝。与两种两阶段目标识别模型（Faster-RCNN和Cascade-RCNN）和单阶段目标检测模型（RetinaNet）相比，该方法在HRNet-40w主干网络结构下实现了79.4%的最高平均精度平均准确率最高可达80.7%。应用阶段平均准确率达到70.8%，召回率达到90.4%。本研究开发的数据集和框架是结合遥感大数据和深度学习方法在全球范围内识别水坝的首次尝试。