Dinâmica da erosão linear em áreas urbanas: Influência das mudanças de cobertura do solo nas áreas de contribuição hídrica utilizando algoritmo de classificação interpretável

Abstract

Linear erosive processes result from various natural factors but are accelerated by anthropogenic factors, such as land use changes. The spatial modeling of erosive processes is widely used in rural contexts to predict losses and prevent damages, and in urban contexts to support territorial planning and management. Currently, the most widespread modeling techniques focus on rural areas, primarily addressing sheet erosion. Studies encompassing urban and peri-urban areas, with a focus on linear erosion, are scarce. In both cases, when incorporated into studies, the land use and land cover factor often fails to consider temporal dynamics of change and is analyzed without accounting for the contributing area to the phenomenon. Acknowledging this gap, the main objective of this thesis was to investigate the effect of land cover transitions within the hydrological contributing area on the triggering of linear erosion in urban and peri-urban areas. To achieve this aim, the research was structured into three articles. Laying the groundwork for this thesis's hypothesis, the first article demonstrated that the variability of the phenomenon is predominantly explained by anthropogenic variables (land use and land cover), which had a greater impact than natural factors. Through a literature review, the second article identified that the sampling of negative instances and the delimitation of their areas of influence emerge as critical methodological aspects for the robustness of machine learning models applied to the analysis of erosive processes. The third article investigated the patterns of land cover change with the greatest influence on the occurrence of erosion. For this purpose, a comparative analysis was conducted, incorporating land use transitions over a twenty-year interval (1990-2010) and adopting two distinct spatial approaches: (i) a 100 m buffer and (ii) the hydrological contributing area of each feature. Decision tree models revealed that the most important transitions for classification were those related to agricultural uses that remained stable or converted to non-vegetated areas. The comparison between approaches demonstrated that the hydrological contributing area was more effective in identifying positive cases of erosion, with the "preserved forest areas" variable being the most relevant in its model. However, the buffer strategy provided a better overall predictive balance between the classes. In summary, the results have the potential to aid in the understanding and prediction of linear erosive processes in urban areas, contributing to the technical-scientific field and to society as a whole by assisting in the prevention of environmental, social, and economic damages.