Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Water Extraction from Tengger Desert Based on Deep Learning

Auteur : 胡李发

Grade : Master 2022

Université : Northwest University

Résumé
Tengger Desert is the fourth largest desert in China.There are many water,but most of them are less than 1 hm 2.Previous studies mostly rely on low-resolution Landsat series,and the traditional object-oriented classification and visual interpretation methods are mainly used in research methods,which are difficult to meet the needs of water body mapping accuracy and efficiency under the current research background.Based on it,this paper proposes a light-weight water body segmentation network with attention mechanism based on the domestic high-resolution images and the full convolution neural network,aiming at the problems of low automation and poor universality of traditional remote sensing water body extraction.Related research results are as follows :（1）Aiming at the problem that large complex networks are difficult to train,a lightweight water segmentation network was designed.Compared with other classical segmentation algorithms,the proposed model has the advantages of small resource consumption,fast reasoning speed,and the best balance between efficiency and accuracy.（2）In order to further improve the generalization ability of the model for multi-classification tasks,a mixed attention module based on group convolution and channel mixing strategy is designed.This module uses two parallel spatial and channel attention units to capture the weights of feature maps at the position and channel levels,respectively.Compared with classical SE and CBAM,SA has obvious advantages in performance.At the same time,the LEDNet-SA is compared with the traditional SVM and water index method.The proposed method is superior to the traditional method in both efficiency and accuracy.（3）In the summer of 2021,there were 1189 water areas with an area of more than 0.01hm 2in the Tengger Desert,and the total area of the water was 2514.17 hm 2.The area and number of water shows a spatial distribution trend of large dispersion and small aggregation.The research shows that the seasonal variation trend of the number and area of water in the region is obvious.The number and area of water reaches the highest value in spring and the lowest value in autumn.Seasonal variation of lake area varied among different lakes,with Aiyite and Abuxige having the smallest annual variation（＜15%）and Tonggulounaoer having the largest annual variation（85.5%）

Mots clés : Tengger desert ;deep learning ;water;lightweight network ;attention mechanism ;

Présentation (CNKI)