Optimizing disturbance mapping in seasonal biomes based on an ensemble framework

Abstract

Mapping and monitoring disturbance in vegetation provide support for developing management strategies, implementing policy initiatives, and providing inputs for modeling ecological and environmental processes. However, seasonal biomes are naturally heterogeneous in terms of climate, soil, biodiversity, and threats posed by human activities and land occupation. In this thesis, mapping and monitoring disturbances in native vegetation were optimized based on ensemble techniques, which uses multiple or committee classifiers combining their predictions. For this purpose, this thesis was organized in three articles. In the first one (1) disturbance maps of seasonal biomes from different spectral indices were evaluated based on the spatial agreement between maps and their accuracies. The results indicated a low rate of spatial agreement among index-based disturbance maps, which was minimally influenced by vegetation domain. In addition, index-based disturbance maps reflected site-specific sensitivity. In the second article (2), the effectiveness of a heterogeneous ensemble classification and data-driven regionalization for improving vegetation disturbance mapping accuracies over large areas was assessed. The ensemble method combined disturbance maps from the LandTrendr algorithm and Random Forest. The results indicated gains in accuracy by the ensemble method compared to non-ensemble methods of disturbance mapping. In addition, data-driven regionalization addressed complexities arising from variability in vegetation types, local climate, and topography across our study area, identifying climate and seasonal metrics as important variables for reducing uncertainties in vegetation disturbance maps. Finally, the third article (3) used object-based image analysis and evaluated predictor variables from both LandTrendr and semivariogram for mapping and characterizing land cover changes. Three classes of land cover changes: non-change, vegetation loss, and pos-change, were set combined with three datasets: LandTrendr, Semivariogram, and Blended. The Blended datasets returned the best accuracies. This article also indicated that semivariogram variables faithfully captured patterns of vegetation loss and recovery. Thus, the increasing need for mapping and monitoring disturbances in seasonal biomes suggests that the methods and algorithms presented in this thesis, return satisfactory accuracies and may be suitable for large-area applications.