Combining UAV and multi-source satellite data for sub-pixel classification of forest vegetation in riparian zones

• Ano de defesa: 2024
• Autor(a) principal: Casagrande, Luan Carlos da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Calibração; Calibration; Machine learning; Satélite; Satellite; Sensors synergy; Sinergia dos sensores; UAV; VANT
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-18122024-145230/

Numerous surface processes can occur in riparian zones and monitoring the spatial-temporal dynamics of land cover is essential to understand their impacts. Deforestation and inadequate use of these areas are notorious problems and can have an important impact on water resources, wildlife, and human communities, among others. Based on its importance, regulations were created aiming to protect these areas. Quickly and accurately mapping forest vegetation near rivers is necessary to guarantee that these regulations are being respected and degraded areas are being recovered. However, mapping riparian zones represents an unique challenge for different reasons, mainly when taking as reference a continental country such as Brazil. Several works were proposed based on unmanned aerial vehicle (UAV) or satellite data and each of these platforms has important advantages and disadvantages. Between these two sources of remote sensing data, the advantages of the former appear to compensate for the disadvantages of the latter. Based on such context, in this work, we proposed to use UAV data to calibrate a multi-source satellite data based model to predict sub-pixel composition in riparian zones. We also perform an analysis aiming to assess the influence of each data type and temporal data as an additional dimension to our research problem. These analyses are performed on a newly compiled dataset composed of data acquired in Europe and South America, which relates UAV data with Sentinel-1 and Sentinel-2 data. Our study shows that, compared to reproduced works, our approach using optimal combinations produced significantly different and superior results for the dominant class. Moreover, it showed promising potential in predicting class membership for pixels containing multiple classes. Besides that, we have demonstrated that the addition of spatial context, Sentinel-1, and spectral indices has helped to improve the results, and that the inclusion of temporal data in the proposed model had a significant impact on the performance, especially in 3D CNNs.