Very high resolution multi-spectral remote sensing image cassifcation

Abstract

As an immediate consequence of the improvements in remote sensing sensors in terms of spatial resolution, and the increase of the amount of information recorded, traditional image segmentation tools became limited in accurately capturing the different land cover objects. This thesis is one more contribution in dealing with the complexity of very high resolution remote sensing image segmentation. We propose an improved version of matched filters, a tool to enhance and extract spatial information in VHR remote sensing images, in the aim to increase the class separability. With the help of particle swarm optimization, we designed an adaptive scheme that tailors matched filters kernels to supposedly meet the requirements of any study area. Traditional matched filters are in fact an out-dated method, mainly due to the choice of the kernels coefficients, which were specifically set for each experiment individually. Moreover, and mainly because of this limitation, matched filters were rarely employed as 2D filters. Our proposed approach automatically generates these kernels, and have shown considerable classification accuracy improvement. Additionally, our work includes various techniques for VHR image segmentation, where we explored textural information, data clustering and deep learning. Furthermore, we conducted a comparative study, contrasting pixel-based (SVM) with object-based (U-Net) image classification methods.