基于深度神经网络的蒙古国色楞格河流域水体信息提取

蒙古高原地处干旱半干旱地区,河流水系对该区域的资源环境格局及其生态环境影响重大。发源于蒙古国的色楞格河是蒙古高原最主要的水资源来源,准确掌握该流域的水体信息对东北亚地区生态环境问题及资源保护具有重要意义。本文以蒙古高原色楞格河流域为研究对象,基于谷歌地球引擎（Google Earth Engine,GEE）云平台,使用 Sentinel-2 多光谱卫星遥感影像,利用深度神经网络（Deep Neural Network, DNN）方法对色楞格河流域的水体信息进行提取,并与支持向量机方法进行对比;利用全球30 m SRTM数据生成水系分布矢量图,通过空间分析形成河流提取目标区,结合深度神经网络分类结果,绘制蒙古国色楞格河流域2019年河流分布图。研究结果表明：① 该方法能够准确地完成大流域范围内的水体制图,提取结果能够体现色楞格河流域河流的空间分布,且能够减少河流断流、空洞现象;② 深度神经网络模型中批量大小设置为8时,在处理数据速度与精度中达到最优,而神经网络结构中隐含层数达到4层时,在精度评价指标测试数据集上达到0.9666,保证了模型特征挖掘能力;③ 经样本点的验证,结果总体精度达到97.65%,可以满足实际应用需求。本研究预期可以为蒙古高原的水体提取提供方法支持和相关数据支持。

Water Information Extraction of Selenga River Basin in Mongolia based on Deep Neural Network

Arid and semi-arid climate zones of the Mongolian Plateau support multiple river systems that significantly impact the distribution of resources and ecological environments of the region. The Selenga River is the most important water resource on the Mongolian Plateau and the only river that enters the largest freshwater lake in the world——Lake Baikal in Russia. Thus, an accurate understanding of the hydrologic characteristics of the Selenga River Basin is critical in monitoring the ecology, environment, and water resources in Northeast Asia. Obtaining the plane morphology of curved rivers in the upper reaches of the plateau region is helpful to analyze the mathematical and physical mechanism of the evolution of curved rivers, which is of great significance to the study of river geomorphology. In this study, the Selenga River Basin in the Mongolia was taken as the study area. Based on the Google Earth Engine (GEE) cloud platform, Sentinel-2 multi-spectral satellite remote sensing images, Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), and automated Water Extraction Index (AWEI) were used to construct a deep neural network model for the extraction of fine water body data in the Selenga River Basin. Three different types of water bodies, including lakes, main streams, and tributaries, were selected to compare the extraction results of deep neural network with the extraction results of support vector machines and Global Surface Water data. Global 30 m SRTM data were then used to generate a river system vector distribution map, identify river extraction target areas, and draw a 10 m spatial resolution river distribution map for the 2019 Selenga River Basin using output from the deep neural network model. Results indicate that, firstly, this method can accurately produce a water regime map for a large basin. The extraction results can reflect the spatial distribution of rivers in the Selenga River Basin in Mongolia, as well as reduce the drying-up and empty phenomenon of rivers. Secondly, the batch size of eight in the deep neural network model optimize data processing speed and accuracy while the four hidden layers in the neural network structure produce an accuracy evaluation index of 0.9666, which guarantees feature mining capabilities of models. Thirdly, after the verification of sample points, the overall accuracy reaches 97.65%, meeting actual application requirements. Thus, this research provides a method and data support for water body extraction for the Mongolian Plateau, which will play an important role in revealing long-term changes in river hydrological conditions, and impact related research on resources and environmental benefits.