Remote Sensing Image Classification Based On Improved Deeplab-v3+Model

The rapid development of remote sensing technology has been widely used in land resources and urban planning.How to process and obtain effective information from a large number of remote sensing images has become the key to the field of remote sensing.Because remote sensing images have the characteristics of high resolution,diverse spectral information,variable target structures,and complex backgrounds,the accuracy of the classification results needs to be further improved in terms of meeting actual needs.In recent years,a large number of scholars have tried to introduce the neural network for RGB three-band true color natural images into the field of remote sensing,and have obtained application effects that are superior to traditional algorithms for classification and target detection.Deep learning methods developed in neural networks can better understand the contour and texture information of objects,automatically learn spatial features and topological relationships from training images,and perform semantic segmentation based on the learned features to obtain higher segmentation accuracy.However,the multi-layer neural network structure of deep learning brings a lot of calculations,and control efficiency is a problem that researchers must face.This paper focuses on the problem of semantic segmentation of remote sensing images based on deep learning.A series of network models represented by fully convolutional neural networks have improved the segmentation accuracy of remote sensing images.Among them,Deeplab-v3+ is currently the best general segmentation network with good segmentation effects and smooth edges,but the training and inference speeds are relatively slow..Based on this,this paper uses Deeplab-v3+ as the basic model,improves the ASPP model in the encoding stage,and uses the network structure combined with Xception-Res Net.In the decoding stage,according to the output of the encoding stage,introduces the layer feature fusion optimization module,and uses CCF competition standard data Comparing and verifying the improved models in the collection,the accuracy and speed of remote sensing image classification are improved,which is 4% higher than the original model accuracy,and the model training speed is increased by 25%.Subsequently,the improved Deeplab-v3+ model was used to analyze the land use distribution characteristics of the Chengdu Ecological Ring.Based on the classification results of remote sensing images,the land use of Chengdu Ecological Ring is mainly cultivated land,construction land,and forest land.Among them,cultivated land and forest land together account for 53.01% of the total area of the ring ecological zone,which is the basis of the ecological corridor.The scale of construction land has been effectively controlled,and the protection of cultivated land has been further valued.However,there are still problems such as poor ecological function,lack of outstanding landscape features,and lack of landscape design and construction.Finally,the study concluded that the use of the improved Deeplab-v3+ model for remote sensing image classification can be used in actual planning.