Abstract
High spatial resolution satellite (HRS) images are being extensively utilized for the detection of changes like urban dynamics, infrastructure surveillance, disaster management, and topographic map-making applications. The enormous information and challenging data are important for the change detection (CD) in these images. However, lack of training data and an excessive amount of information, which discourage the researcher from developing a deep-learning-based efficient algorithm for CD. Therefore we want to develop an efficient algorithm in terms of accuracy and speed. Hence, we gain attention on designing an optimized Convolution Neural Network (CNN) while maintaining the speed and segmentation accuracy of the network. We develop an optimized architecture called as EffCDNet which adopts a siamese-based pre-trained encoder with an Attention-based UNet decoder for semantic segmentation. The network is built with pre-trained EfficientNet architecture with shared weights for strong feature extraction and to overcome the limitations caused by insufficient training data. The attention-based UNet decoder uses the attention-gate layer mechanism right before concatenation operation. This obtains more discriminative relevant features for improving the segmentation performance. Also, it is used for the reconstruction of fine-grained feature maps with significant use of context information. For improvement in the change map, we used the Undecimated Discrete Wavelet Transform (UDWT) fusion as a post-processing technique for spatial and temporal analysis of multi-resolution images to obtain a much more enhanced information difference map. The resulting image is less affected by noise, shift-invariable, and overcomes the mixed pixel problem to detect small possible changes. Experimental results on LEVIR-CD, SZATKI AirChange (AC), and Onera Satellite Change Detection (OSCD) benchmark datasets proved that the proposed approach outperforms its superiority in terms of Intersection over Union (IoU) and inference time over the existing methods.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.