Research On Change Detection In Remote Sensing Image Based On Deep Learning

Remote sensing image change detection is one of the important research directions in the field of remote sensing image processing,and it is also one of the main driving forces to promote the continuous development of remote sensing image processing technology.Due to the rapid development of urbanization and various human activities,the earth’s surface is constantly changing.Such changes have a profound impact on the earth’s environment and resources.Therefore,timely and effective observation of the earth’s surface changes is very necessary.At the same time,with the development of aerospace technology,remote sensing image data is becoming more and more accessible,and the resolution of remote sensing images is getting higher and higher.High-resolution remote sensing images mean that they contain richer and more detailed image information.It also provides a sufficient source of data for monitoring changes in the earth’s surface.The traditional change detection method relies more on the experience and knowledge of remote sensing interpreters,and also requires rigorous and complex mathematical background knowledge when analyzing images,which wastes manpower and material resources during interpretation.Therefore,it is very meaningful to design a flexible automatic change detection method for the field of remote sensing image processing.Recently,with the rapid development of artificial intelligence technology and computer equipment,many methods based on supervised convolutional neural networks have been proposed for change detection tasks.However,the current change detection method still has some shortcomings.One is the problem that small changing targets are prone to miss detection during detection,and the other is that the edges of changing targets are prone to error detection.The research of this paper mainly starts from the above problems.According to the existing change detection architecture idea,a new method of change detection based on deep learning is designed.This method mixes deep learning and existing architectural ideas.The method in this paper is based on the current commonly used encoder and decoder architectures,and the different change detection method is designed with three different encoder ideas.In order to solve the problem of poor detection effect of small change targets and inaccurate edge detection of changing targets.From the perspective of feature reuse,by making full-scale skip connections between encoder features and decoder features at different levels,shallow low-level information and deep high-level semantics can be fully utilized,and full-scale skip connections can alleviate the loss of feature spatial details.In this way,the detection ability of small changing targets has been improved.From the perspective of feature selection,a multi-receptive field position information enhancement module is designed based on coordinate attention,which highlights effective features,suppresses invalid noise features,and enhances the spatial position relationship of the encoder output features at all levels.This can more accurately detect the edges of changing targets.Finally,this paper also introduces depthwise over-parameterized convolutional layer(DOConv)in the designed method to replace the traditional convolution operation.DOConv improves the feature extraction ability of the network without increasing the complexity of inference calculation.In order to verify the effectiveness of the proposed method,this paper has been verified on some public data set.The experimental results show that the proposed method has good change detection performance.