Abstract :
This paper introduces a boundary detection technique for remotely sensed images using a CNN-based multi-band fusion architecture. Conventional edge detection techniques fail to cope with the improvements in spatial, spectral, and radiometric resolutions of remote sensing images. While current approaches have handled each complexity in a mutually exclusive manner through specific adaptation of boundary detection parameters, there have been limited techniques that are feature-independent and parameter-free. The proposed approach attempts to integrate complementary and redundant information from the various multi-spectral bands of remotely sensed images to provide a composite image which could enhance perception and reinforce common interpretation while facilitating self-learning and customization capabilities for detection and fusion respectively. First, the technique associates a confidence map on the location of region boundaries with each multi-spectral band individually using a Convolutional Neural Network (CNN). Further, a weighted decision-based fusion framework is applied to integrate the contributions of individual confidence maps into one unified boundary detected output. Systematic experiments are conducted on publicly available datasets in order to evaluate the performance of the method and benchmark it against competing baselines.
