Research On Change Detection Of High-resolution Remote Sensing Images Based On Semantic Information Enhancemen

Traditional change detection methods,while theoretically and practically simple,compute shallow features on a per-pixel basis,ignoring spatial contextual information and producing a lot of salt-and-pepper noise,making it difficult to accurately identify changes in land cover types between images.Additionally,traditional change detection models built from a specific area of remote sensing images have poor generalization capabilities and cannot perform automated change detection,making them unsuitable for large-scale change detection tasks.In recent years,with the maturation of computer technology,deep learning has provided more powerful technical support for remote sensing image change detection.Change detection methods based on deep learning automatically extract deep change features from remote sensing images,segment remote sensing images,and thus greatly reduce manual feature engineering,while completing large-scale change detection tasks more robustly.Convolutional neural networks(CNNs)have been successfully used in change detection algorithms with strong discriminative power.CNN-based methods convert two temporal images into high-level spatial and depth features,thus minimizing error propagation and artificial errors caused by preprocessing.To address the challenges posed by complex image background information interference and difficulties in capturing small-scale land cover changes under different lighting conditions,this paper proposes a change detection model based on FCNN-DCRF.The model consists of an encoder and a corresponding decoder network and does not use fully connected layers.In the FCNN-DCRF decoding network,the decoder receives the image feature map output from the corresponding encoding network and completes the upsampling operation on the feature map through maximum pooling indices,outputting rough feature extraction results.Then,the upsampled rough features are filtered and convolved in the decoding network to complete the precise extraction of change detection features.Finally,the decoding network outputs the change result map through the Softmax function to achieve pixel-by-pixel change detection.Although FCNN-DCRF has achieved good results,the limited receptive field of CNNs has limitations,and the introduction of fully connected conditional random fields has to some extent improved the model’s ability to extract global semantic information,but it still needs improvement in identifying pseudo-changes and refining the boundaries of changing areas.Therefore,based on the latest research progress in deep learning,this paper proposes an LGENet,an end-to-end encoding-decoding network for local-global feature-enhanced Transformer change detection model.The model uses the LE-Transformer module to expand the model’s receptive field,enhance its ability to extract global features,improve the continuity of the model’s prediction results,the ability to detect small targets,and obtain more detailed detection area boundaries,while eliminating pseudo-change interference caused by various complex lighting and seasonal conditions.Comprehensive comparative experiments on the LEVIR-CD and CDD change detection datasets show that the FCNN-DCRF model proposed in this paper achieves the highest precision,F1 score,and Io U values on the LEVIR-CD dataset,reaching89.65%,87.85%,and 81.49%,respectively.This shows that embedding the DCRF model in the fully convolutional feature extraction network can indeed more accurately identify changing areas and obtain more accurate change detection results.