Development of land use mapping methods based on multi-sensor data and machine learning algorithms

Abstract

Different agricultural mapping initiatives have been carried out to support rural planning and crop prediction. However, due to the complex dynamics and scale of Brazilian agriculture, there is a need to generate methodologies with greater detail, frequency and precision to discriminate between agricultural crops, pastures and forestry. The increase in spatial, temporal and spectral resolution of available satellite images, the high computational capacity for processing large volumes of data, and the development of artificial intelligence associated with classifiers based on machine learning algorithms, enhance research in agricultural remote sensing control for classification, monitoring and agricultural management. However, generating more effective and accurate multi-sensor agricultural mapping still presents a major challenge, due to: i) the level of preprocessing required in different images and its relationship with the integrity and homogeneity of agricultural targets; ii) the complexity of combining different temporal and spatial resolutions, which allows daily analyzes of less than 0.5 meters; iii) due to the difficulty in generating multi-sensor analyzes that enhance digital classification methods via machine learning. With the implementation of this post-doctoral scholarship, the objective is to develop agricultural mapping methods based on machine learning algorithms and a multi-sensor approach with images from satellites, nanosatellites and UAVs in the regions of the Agrotechnological Districts (DAT's) defined by the Center. Among the results, it is expected to generate new methodologies applied to agricultural monitoring, qualified scientific publications, and promote postgraduate professional training.