Gap filling of optical remote sensing multi-source data cube through multi-scale and multi-temporal segmentation

• Ano de defesa: 2019
• Autor(a) principal: Rennan de Freitas Bezerra Marujo
• Orientador(a): Leila Maria Garcia Fonseca
• Banca de defesa: Sidnei João Siqueira Sant'Anna, Thales Sehn Körting, Giovanni de Araujo Boggione, Silvia Helena Modenese Gorla da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Instituto Nacional de Pesquisas Espaciais (INPE)
• Programa de Pós-Graduação: Programa de Pós-Graduação do INPE em Computação Aplicada
• País: BR
• Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21c/2019/03.20.17.01

A promising solution to solve the lack of Earths surface observation problem on multi-spectral images consists in integrating multi-sensor data. However data must be harmonized before its measures can be comparable and a treatment for gaps (due to cloud cover, sensor defects, and partial images) must be considered. In this context, the present research proposes a methodology to build a gap-free multi-source and multi-spectral data cube, which involves Earth Observation data harmonization and reconstruction. To accomplish this, we tested two harmonization procedures, one based on linear regression and the other based on linear unmixing model, and propose a procedure for spatial-temporal gap-filling, which does not require previous reference. Two approaches for filling gaps are developed. The first one aims at improving a method based on spatial context of close-in-time images to fill small clouds and stripe effects, where in our adaptation a weighting factor is used for each pixel within segments. The second one uses a multi-temporal segmentation to fill the remaining gaps. The gap-filling strategies are applied on two image data cubes composed by Landsat-7/ETM+, Landsat-8/OLI images and CBERS-4/MUX images. To validate the gap-filling procedure, we simulate artificial gaps in the images and, subsequently we compare the original image with the gap-filled ones. Our approach based on weighting factor surpassed the original method for all bands, presenting R2 greater than 0.90 and a V IF of at least 0.97, while asymptotically maintaining the algorithm cost. It also preserved the texture on reconstructed images, and also was capable of detecting narrow features, e.g., roads, riparian areas, and small streams. The second approach based on multi-temporal segmentation filled all the remaining gaps, 43.64% of the entire data cube. However, the estimated values are more affected by uncertainty and the image texture is affected, resulting in a homogeneous gap-filling. The harmonized and reconstructed areas were very similar to the original data, presenting an UIQI of at least 0.92 and a V IF ranging from 0.6 to 0.7 on the final method, showing the feasibility of the methodology.