Classificação da cobertura do solo com base em imagens RGB e multiespectrais obtidas por VANT: estudo de caso para o bioma caatinga

Abstract

The Brazilian Caatinga is a biome that corresponds to one of the largest Seasonally Dry Tropical Forests (SDTF) in the world and understanding the land cover in this area contributes to studies aimed at environmental preservation and mitigation of environmental impacts. One of the challenges in classifying land cover in this area is the distinction between different vegetative patterns. In this context, Unmanned Aerial Vehicles (UAVs) emerge, capable of obtaining data with high spatial resolution and overcoming this limitation. This research evaluates the accuracy of land cover classification using data obtained from a UAV with an onboard multispectral camera. Two approaches using the Random Forest (RF) classifier were applied, the first based on the Red, Green, Nir and NDVI bands obtained from the multispectral camera and the other based on the Red, Green, Blue bands and NGRDI obtained from the RGB camera's native UAV in flights carried out at 120 meters high with clear sky conditions and centered sun, in the Riacho do Frango Basin in Patos-PB. The collection period was from 09/22/2023 to 10/22/2023, fully included in the dry season. The classification results were evaluated based on model validation samples and indicate that the first approach presented the best performance with Global Accuracy of 95.33% and Kappa Coefficient of 93.33%. The most important variable, according to the MDG index, was the NDVI, which shows the importance of using vegetation indices to highlight vegetative characteristics. From the results, it is concluded that the use of UAV with the onboard Mapir Survey 3w camera is an efficient tool for classifying land cover in the Caatinga, with a high capacity for distinguishing vegetative patterns, while the classification for the area under study made based on RGB data alone does not provide good accuracy.