Gestão hídrica e desenvolvimento rural sustentável: desenvolvimento de um sistema utilizando georreferenciamento e inteligência artificial para prevenção de erosão em mananciais de abastecimento público

Abstract

The intensification of agricultural land use in rural watersheds has increased erosion processes and compromised the quality and availability of water resources, especially in areas dominated by family farming. The aim of this thesis is to develop and validate an integrated methodology based on aerial photogrammetry using Remotely Piloted Aircraft Systems (RPAS), geoprocessing, derived attributes calculated from RGB orthomosaics, artificial intelligence, and independent validation using the Revised Universal Soil Loss Equation (RUSLE), with the goal of identifying areas susceptible to erosion, simulating environmental restoration scenarios, and informing strategies for water management and sustainable rural development. The study was conducted in a 59.24-hectare area located in one of the headwaters of Arroio Fundo, a public water source for the municipality of Marechal Cândido Rondon (PR), situated on the border with the municipality of Quatro Pontes, in western Paraná. The methodology involved the acquisition of multitemporal aerial images via UAV with a spatial resolution of 3.7 cm/pixel, the development of an automated computational pipeline for photogrammetric processing, the generation of derived attributes, the preparation of data for training the neural network, and the semantic classification of land use and land cover. For land use and land cover classification, a U-Net-type convolutional neural network was used, optimized through data augmentation and class weighting strategies, and subsequently integrated with topographic variables to model erosion risk using map algebra. The consistency of the results was verified using the RUSLE. The results of the current scenario assessment (2025) showed that 60.31% of the study area is at high risk of erosion, with 1.61 hectares of active erosion processes identified, for which the average soil loss estimated by RUSLE reached 114.52 t ha⁻¹ year⁻¹. The simulation of the prospective scenario (2035) indicated that the full restoration of Permanent Preservation Areas could completely eliminate areas at critical risk and reduce high-risk areas by 72.8%, with a 319.9% increase in low-risk areas. The proposed integrated methodology is a technical tool with low operational costs, based primarily on free and open-source software, with the potential to be replicated by municipal technical teams and research institutions in similar rural contexts, contributing to territorial planning geared toward sustainable rural development.