Use of dense Sentinel-2/MSI time series and machine learning algorithms to improve crop monitoring in the Cerrado biome

Abstract

Endowed with unique natural resources and agricultural production, the Cerrado is a strategic biome for Brazil. However, the dichotomy between conservation and production raises concerns associated with land use and land cover changes. Agricultural production is intense, especially in the MATOPIBA, a territorial geographic reality rich in edaphoclimatic characteristics and technology for high yields. However, as the area and volumes of production and exports grow, the biodiversity loss intensifies climate change and generates economic consequences. With changes in the global consumption profile, less tolerant to products derived from illegally deforested areas, Brazilian exports have suffered boycott threats in the international market. The agricultural sector has been considered a villain of deforestation and needs to improve the monitoring and traceability of the supply chain. Precise land use and land cover information subsidize it and climate change, food security, and agricultural dynamics policies, among others. However, solving it still is a challenge. Due to data scarcity or models disconnected from the field's reality, accurate analyzes only occur in the post-harvest period, inhibiting the detailed and updated change detection. This proposal aims to develop a method to improve agricultural mapping in the Cerrado from machine learning and time series analysis derived from analysis-ready data composed of data from the Sentinel-2/MultiSpectral Instrument mission, sensitive to subtle variations in vegetation, in a multidimensional data cube architecture in space, time, and properties. It is expected to propose a within-season agricultural monitoring system and generate accurate information at the level of crops and natural vegetation phytophysiognomies. (AU)