Research On Lake Extraction Of Qinghai-Tibet Plateau Based On PIE-Engine Platform

As an important component of the Asian Water Tower,the Qinghai-Tibet Plateau is rich in water resources.As an indicator of climate change,widely distributed lakes occupy an important role in the local water cycle.With the increase of the resolution of remote sensing images,people’s demand for the accuracy and efficiency of lake extraction is growing.Therefore,accurate and efficient extraction of lakes in the Qinghai-Tibet Plateau is of great significance in the fields of water resources,climate change and disaster prevention.In order to achieve the accurate and efficient extraction of lakes,this study based on Sentinel-2 images to build a semantic segmentation data set,and designed the extraction process of lakes on the Qinghai-Tibet Plateau that combines deep learning with PIE-Engine cloud platform.By comparing the common semantic segmentation model and the water index method,the advantage of the deep learning network architecture proposed in this paper in the lake extraction task is verified.In addition,based on Sentinel-2 images,the data set of lakes with an area of over than 0.01 km~2 in the Qinghai-Tibet Plateau in China in 2022 is obtained,and the reliability of the extracted data set of this study is verified based on the Qinghai-Tibet Plateau lake data set released by Zhang et al..The main conclusions of this study are as follows.(1)Based on Sentinel-2 MSI image data,this study create a sample dataset for lake extraction.After clipping and expansion,9075 sample images are generated with the size of 512 by 512 by 4.The training results show that the sample data can support the training task of the lake extraction model.(2)We proposed a new attention mechanism and improves the Deeplab V3 plus baseline network by combining with the MobileNet V2 lightweight network.In local comparative experiments with common semantic segmentation network models,and common water index methods(NDWI,MNDWI),the method presented in this paper exhibits better advantages,with a local test MIoU of 92.46%.(3)Upload the network model to the PIE-Engine AI platform to implement online lake recognition model training.The visual intermediate training results show that the model has effectively learned lake features.According to the dynamic loss function curve provided by the platform,the optimal model was achieved in 193 rounds of training,with a precision of 98.02%and a recall rate of 95.69%.(4)Based on the PIE-Engine Studio cloud platform,more than 2900 Sentinel-2 images in China on the Qinghai-Tibet Plateau are called,and the images are fused in the way of median,and finally uploaded to the image intelligent interpretation platform in batches to complete the interpretation application combined with the trained model.The area regression statistics of the lake data extracted in this study and the dataset released by Zhang et al.show that R~2 is 0.9998,which indicates that the lake extracted in this study has certain reliability.(5)In 2022,the Qinghai-Tibet Plateau has 36160 lakes with an area of more than 0.01 km~2 in China,with a total area of 52244.05 km~2.In the study area,the number of lakes distributed in the Inner sub-basin is the largest(19246)and the total area is the largest(36385.85 km~2),while the number of lakes distributed in the Mekong sub-basin is the smallest(134)and the total area is the smallest(22.57 km~2).