The geostatistical context employed in remote sensing applications: image classification, change detection and forest inventory

Abstract

We used the spatial context, specifically the geostatistical techniques to improve remote sensing applications: image classification, change detection and forest inventory. We first evaluate the potential of semivariogram parameters, derived from satellite images with different spatial resolutions to characterize landscape spatial heterogeneity. In the image classification study, the goal was to assess the potential of geostatistical features at the object level to improve the image classification of contrasted landscape vegetation cover. In the change detection approaches we explored and evaluated the performance of semivariogram indices in an object-based approach to detecting land-cover changes using the NDVI derived from Landsat images using the support vector machines and random forest algorithms. We assessed the potential of geostatistical features to accurately detect land-cover changes, disregarding those associated with phenological differences. In the forest inventory manuscript, we investigated the potential of data extracted from Landsat TM, MODIS products and spatialenvironmental variables to map the spatial distribution of aboveground biomass in Minas Gerais State, using random forest regression algorithm and regression kriging technique using a stratified design. The applications results indicate that: (1) image spatial resolution does in fact influence the sill and range parameters, that can be used as a simple indicator of landscape heterogeneity; (2) semivariogram curves were efficient for characterizing spatial heterogeneity, significantly improving the image classification accuracy when combining geostatistical features with spectral data; (3) geostatistical features have the potential to discriminate between homogeneous and heterogeneous classes within objects, are not affected by vegetation seasonality, and can produce times series that accurately differentiate forest changes from seasonal changes, resulting in fewer classification errors and (4) the stratification of data into vegetation types not only improved the accuracy of aboveground biomass estimative, but also allowed random forest regression to select the lowest number of variables that offer the best predictive model performance to AGB mapping. The spatial context approach we presented in this thesis is a novel and useful remote sensing method for the image classification of spectrally similar land-cover types, detection of forest change events in areas where forests exhibit strong seasonality, and therefining aboveground biomass map and the understanding of how the variables properties are associated with the biomass enable researches to improve the roughly estimates of greenhouse gas emission and also helps the selection of appropriate variables that best model the aboveground biomass in savanna-forest transition areas.