Mapeamento de uso da terra e cobertura vegetal do bioma caatinga do Rio Grande do Norte: estratégia baseada em aprendizado de máquina e computação em nuvem

Abstract

Landscape transformations resulting from the replacement of native forests by human activities have generated significant impacts on natural areas. Mapping land use and cover is essential to support territorial planning, since changes in Caatinga vegetation in semi-arid environments are associated with rainfall variability and anthropogenic action. In view of this, this research aimed to map land use and land cover in the Caatinga biome in the state of Rio Grande do Norte for the year 2023, using the Random Forest algorithm and cloud computing. In addition, we sought to analyze the relevance of environmental variables, especially climatic components, in the spatial behavior of vegetation patterns. The methodological approach involved two main stages: spatial prediction of precipitation and supervised classification, both using machine learning algorithms. The LANDSAT-8 image collection was processed using the Google Earth Engine (GEE) platform. The modeling used environmental variables such as estimated precipitation, surface temperature, morphometric parameters and biophysical indices. The Feature Importance feature was used to identify the relevance of the 153 covariates. The results revealed an overall average accuracy corresponding to 83% of the real data. The spatialization of forest vegetation classes and water bodies was largely consistent. Imprecisions were detected between Urban Areas and Other Uses, and between Savannah Vegetation, Agriculture and Pasture. The most important variables were spatial position, altitude and climate. The NIR and SWIR bands showed a strong influence, as did the NDTI vegetation index, transformed bands and fraction images. However, the NDVI, GNDVI, EVI and SAVI indices were of low relevance. This research explored advanced satellite data processing and analysis techniques in order to obtain a classification model with good accuracy and high reliability for spatializing the biophysical heterogeneity of the Caatinga. The approach was successful, resulting in positive statistical evaluations and satisfactory spatialization of the 12 grouped classes.