October 07-10, 2018, Athens, Greece | Rodrigo Caye Daudt1,2, Bertrand Le Saux1, Alexandre Boulch1
This paper introduces three fully convolutional neural network (FCNN) architectures designed for change detection using pairs of coregistered images. The authors propose two Siamese extensions of fully convolutional networks, which incorporate heuristics to achieve superior performance on two open datasets, both in RGB and multispectral images. These architectures are trained from scratch using annotated change detection images and outperform previous methods while being at least 500 times faster. The work is a significant step towards efficient processing of large-scale Earth observation data from systems like Copernicus and Landsat. The paper includes a detailed description of the proposed architectures, experimental results, and comparisons with existing methods, highlighting the effectiveness and efficiency of the proposed approaches.This paper introduces three fully convolutional neural network (FCNN) architectures designed for change detection using pairs of coregistered images. The authors propose two Siamese extensions of fully convolutional networks, which incorporate heuristics to achieve superior performance on two open datasets, both in RGB and multispectral images. These architectures are trained from scratch using annotated change detection images and outperform previous methods while being at least 500 times faster. The work is a significant step towards efficient processing of large-scale Earth observation data from systems like Copernicus and Landsat. The paper includes a detailed description of the proposed architectures, experimental results, and comparisons with existing methods, highlighting the effectiveness and efficiency of the proposed approaches.