A high-resolution remote sensing images change detection method via the integration of dense connections and self-attention mechanisms
Remote sensing image change detection is an important task in remote sensing image analysis,which is widely used in urban dynamic monitoring,geographic information update,natural disaster monitoring,illegal building investigation,military target strike effect analysis,and land and resources survey.As a pixel-level prediction task,the current methods of change detection have two prominent problems:one is the computational efficiency of self-attention between arbitrary pixel pairs is low,and the long context information in remote sensing images is insufficiently utilized;the other is that the current methods focus on the extraction of deep change image features while shallow information containing high-resolution and fine-grained features are ignored.To address the first problem,Transformer is used to perform context modeling on the extracted bitemporal image features to improve the quality of the deepest change features.To take into account the efficiency of the Transformer,the proposed method converts the images into sparse tokens,thereby significantly reducing the number of tokens of the Transformer.For the second problem,the proposed method uses dense skip connections to retain high resolution in shallow change features.Three publicly available datasets were used for experiments.Extensive experiments show that com-pared with the state-of-the-art change detection methods,the IoU metric of the proposed method reached 85.44％,84.15％and 94.61％,respectively,which is better than other comparison methods.
万方数据
维普
