Stochastic simulation to improve land-cover estimates derived from coarse spatial resolution satellite imagery

Today's land-cover monitoring studies at regional to global scales using optical satellite based remote sensing is confined to the use of coarse spatial resolution imagery. Due to the coarse spatial resolution, land-cover identification is poor and estimation is error prone. Some land-cover investigations apply a scaling-up approach where fine spatial resolution imagery is aggregated until the wanted mapping scale is attained. Here, the opposite approach (scaling-down) is investigated through the use of geostatistical stochastic imaging techniques. The objective of this thesis is to examine the possibility of generating finer spatial resolution multi-spectral like images based on available multi-spectral coarse spatial resolution imagery to extract land-cover information. Other ideas addressed were: (a) whether stochastic imaging can indeed generate multi-spectral like finer spatial resolution imagery based on coarse spatial resolution imagery, (b) the possibility of introducing SAR imagery to improve the spatial location of the generated image features and, (c) the applicability of an automatic spectral segmentation algorithm to the generated imagery. The sequential gaussian simulation algorithm was used to generate the finer spatial resolution multi-spectral like images. This algorithm was applied using the local varying mean and co-simulation options and always conditioned to the coarse spatial resolution imagery. From aboard the SPOT-4 satellite, the VEGETATION (VGT) instrument provided the coarse spatial resolution imagery while the HRVIR instrument provided the fine spatial resolution imagery for validation. Algorithm parameters were taken directly and derived from the VGT imagery. RADARSAT ScanSAR wide imagery was also used in the stochastic imaging process. Four test sites in the vicinity of the Island of Montreal were chosen each measuring 15 km x 15 km. The tests resulted in the generation of multi-spectral images with red, NIR and SWIR bands. Three different sets of input parameters were used to generate the finer spatial resolution images. The first set was based on the VGT image statistics, the second was based on derived finer spatial resolution statistics and the last set was based on the second set of parameters with the inclusion of SAR imagery. The K-means algorithm was chosen to segment the generated finer spatial resolution images. Overall, this experiment served to: (a) demonstrate that coarse spatial resolution imagery can be applied to generation of finer spatial resolution imagery with stochastic imaging techniques. However, before spectral reproducibility can be achieved, the sensing system and scale relationships must be better understood; (b) illustrate the appropriateness of the co-simulation technique but also show that the input parameters (variogram and distribution) have a significant impact on the resulting scale of the generated finer spatial resolution images; (c) demonstrate that the use of SAR imagery is beneficial to the process of generating finer spatial resolution imagery because it helps fix the ground scene characteristics, but the relationship to the optical imagery (an important input parameter for co-simulation) varies depending on the scene and must be further investigated; (d) show that spectral segmentation of synthetic imagery is possible but validation remains difficult using the standard approach.