Research On Land Classification Method Of Remote Sensing Image Based On Deep Learning

Efficient and accurate ground cover classification using remote sensing images is of great significance in practical applications such as environmental protection,urban planning,land use,and natural disaster prevention and control.In recent years,thanks to the powerful feature induction learning capability,deep learning methods represented by CNN have introduced too many parameters that hinder the deployment of models for applications,although they substantially surpass traditional methods in terms of accuracy.On the other hand,the universality and segmentation accuracy of existing deep learning methods are still not high enough under the influence of factors such as multi-scale,data quality,and model performance,and high-precision feature coverage classification still faces many challenges.Based on this,this thesis conducts an in-depth study on the shortcomings of deep learning technology in remote sensing image ground cover classification research,and the main research works and contents are as follows:1.This thesis proposes a land-cover classification network based on attention enhancement and dense multi-scale features,aiming to address the issue of insufficient segmentation detail caused by multi-scale variations and model performance.This model can better separate spectral similarity features,effectively improving the accuracy and reliability of segmentation.First,in the encoder stage,an attention-enhanced dense cavity pyramidal pooling structure is designed based on the idea of dense network to expand the perceptual field of the model,while a dual attention mechanism is introduced after the average pooling layer to help remote sensing images balance between feature representation capability and spatial localization accuracy.Dense connectivity is employed in the encoder stage to make full use of the multi-scale information of the images.Finally,the comparison experiments with other network models verify that the proposed method outperforms the same type or mainstream remote sensing image feature coverage classification models in terms of accuracy and visualization effects.2.To address the slow inference speed of deep learning in remote sensing image land cover classification,this thesis proposes a lightweight land cover classification network model based on multi-scale and edge-aware features.In the encoder stage,the model uses Res Net18,a lightweight classification network that takes into account the feature extraction effect,as the backbone network to extract image features.In the decoder stage,the method makes full use of the semantic information of the low-resolution feature maps and the spatial information of the high-resolution feature maps.It also assists supervised semantic segmentation master task learning by means of edge-aware loss functions.In this way,the model not only makes full use of the semantic,spatial and edge information of the images,but also circumvents its drawbacks of slow speed and high memory occupation during inference.The experimental results show that the method achieves a balance between recognition accuracy and inference speed,and can meet the demand of real-time feature classification of remote sensing images.3.Using the PyQt5 platform,the proposed land cover classification method based on remote sensing image and the methods used in comparative experiments are integrated to realize automation and visualization of land cover classification for remote sensing imagery.