Research On Semantic Segmentation And Change Detection Methods For High-resolution Remote Sensing Images Based On Deep Learning

With the rapid development of remote sensing aerial observation technology and improvement in the application of imaging sensors,a large number of remote sensing images with higher spatial and spectral resolutions are available for use.The imaging quality of remote sensing images has greatly improved,and obtaining different types of remote sensing data has become more convenient.Remote sensing data can provide an important data foundation for applications such as land resource surveys,urban planning,environmental protection,disaster and agriculture monitoring,etc.It is crucial to use remote sensing images to detect land surface cover types for understanding surface dynamics.High-resolution remote sensing images contain rich ground features,but noise interference is also significant.Images contain more complex ground information,and the phenomenon of different objects with the same spectrum or the same objects with different spectra is more common,making it difficult for traditional methods to achieve satisfactory results.Currently,deep learning technology is developing rapidly and has a natural advantage in image vision.It has excellent feature extraction and expression capabilities compared to traditional pixel-based and object-oriented methods,maintaining high recognition accuracy.Supported by massive and diverse data sources,intelligent interpretation based on remote sensing imagery is an important direction for research on land cover semantic segmentation and change detection.As a new type of technology,deep learning can be used as a new tool for recognizing remote sensing image information.However,in practical applications,existing deep learning methods may suffer from insufficient feature learning ability,single feature fusion mode,and redundant feature information,resulting in rough segmentation results.In this paper,by analyzing the strengths and weaknesses of existing deep learning methods,combined with ideas such as multi-feature fusion and attention mechanism,we research and design optimization modules,build improved deep learning models,and improve the accuracy of remote sensing image semantic segmentation and change detection.This allows for a better realization of the potential application value of deep learning models in remote sensing imagery.The main research content is as follows:(1)To improve the global feature modeling ability of semantic segmentation network,a semantic segmentation model combining convolutional neural networks(CNN)with Transformer is proposed.Firstly,multi-level features are extracted from remote sensing images by the CNN encoder,and the feature pyramid structure is used to fuse hierarchical features,making the fused features contain multi-level feature information.Secondly,the CNNTransformer structure is used to extract global-local features of the fused features to fully extract the feature information of the image.Finally,the attention feature fusion module is used to fuse the features from the CNN-Transformer with the top features of the feature pyramid,further reducing the redundant information and improving the accuracy of the model.Experimental results show that the proposed method can achieve high-precision land cover semantic segmentation results in complex scenarios,with strong robustness and generalization ability.(2)To address the problem of single-scale feature fusion and feature redundancy in common deep learning change detection methods,an attention-guided full-scale feature aggregation change detection network is proposed.Firstly,the full-scale feature aggregation structure is used to fuse the multi-scale features of the encoder,which enhances the relationship between features at multiple scales and reduces the impact of encoder downsampling on the loss of details.Secondly,the attention mechanism is used to refine the fused features by filtering out redundant and irrelevant feature information.Finally,the multiple side-outputs fusion structure is used to fuse the prediction results at multiple scales,making the model adaptable to change detection in different scale regions.Experimental results show that the proposed method outperforms other mainstream change detection methods in detection accuracy and has a good balance between accuracy and complexity.