Leveraging the application of Earth observation data for mapping and monitoring cropland soils

• Ano de defesa: 2020
• Autor(a) principal: Safanelli, José Lucas
• Tipo de documento: Tese
• 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; Digital soil mapping; Machine learning; Mapeamento digital do solo; Pedometria; Pedometrics; Remote sensing; Sensoriamento remoto
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11140/tde-26112020-163718/

The use and sustainable development of cropland soils requires the continuous monitoring and promotion of good practices that support soil quality and the provision of its several functions. As the soil quality and functioning can be affected by several factors and interventions, resulting in changes at the temporal and spatial scales, Earth observation (EO) systems become an sound alternative for monitoring soils due to ability in providing data in a timely manner, covering large geographical areas, and revisiting the same place in Earth in short periods of time. Furthermore, as the availability of detailed information about cropland soils is still a challenge in most countries, and recent literature has been supporting the proposition that collections of EO data are a valuable source for environmental studies, this study aimed at exploring the collection of satellite images for mapping and monitoring cropland soils over large geographical areas. For this, we developed the routines for processing big EO data within a high-performance cloud-based platform. With the combination of extracted features from EO data, legacy soil datasets, and machine learning algorithms, we performed the medium-resolution mapping of cropland soils over the geographical extents of Europe and Brazil. We demonstrated in this study that the collection of Landsat images is a valuable source for extracting spectral features useful for mapping and monitoring cropland soils. The bare soil composite based on the median of 37 years of Landsat imagery allowed the prediction of clay and calcium carbonates with moderate performance in Europe. In addition to that, using the Google Earth Engine, we developed and made publicly available a package to calculate terrain attributes customized to different spatial resolutions, which can be scaled up to the global extent. This package was particularly important for preparing additional information for mapping the cropland soils in Brazil. The spectral and terrain features extracted from EO data allowed the calibration of prediction models of clay, sand, soil organic carbon (SOC) content, and SOC stock with satisfactory accuracy across the Brazilian cropland soils. With the resulting maps, we were able to estimate the total SOC stock and identify some aspects related to the distribution of soil attributes regarding the main agricultural regions. Therefore, this study supports the proposition that EO data is a valuable source for extracting environmental features for mapping and monitoring cropland soils at finer resolutions, assisting the evaluation of soil spatial distribution and the historical agriculture expansion in Europe and Brazil.