Classificação da cobertura do solo da caatinga: avaliação de abordagens em Machine Learning utilizando dados LandSat Mono e multitemporais.

Abstract

Accurate information on the land cover is crucial for efficient monitoring and development of environmental studies in the Brazilian Caatinga forest, one of the largest and most biodiverse dry forests on the planet. Distinguishing different patterns of land cover through medium spatial­resolution remote sensing, such as the Landsat image series, is challenging to Caatinga due to heterogeneous land cover, complex climate­soil ­ vegetation interactions, and anthropogenic disturbance. Two remote sensing approaches have a high potential for accurate and efficient land­cover mapping in Caatinga: single and multi­date imagery. The heterogeneity of the land cover of this environment can contribute to a better performance of multispectral approaches that are usually applied for single­date images. In a land­cover mapping effort in Caatinga, the temporal factor gains relevance, and the use of time series can bring advantages, but, in general, this approach uses vegetation index, losing multispectral information. This manuscript assesses the accuracies and advantages of single­date multispectral and multi­date Normalized Difference Vegetation Index (NDVI) approaches in land­cover classification. Both approaches use the Random Forest method, and the results are evaluated based on samples collected during field surveys. Results indicate that land­cover classification obtained from multi­date NDVI performs better (overall accuracy of 88.8% and kappa of 0.86) than single­date multispectral data (overall accuracy of 81.4% and kappa coefficient of 0.78). The Z­test indicated that the difference in performance between the two approaches was statistically significant. The lower performance observed for single­date multispectral classification is due to similarities in spectral responses for targets of deciduous vegetation that lose their foliage and can be misread as non­vegetated areas. Meanwhile, an accurate classification by time series of plant clusters in seasonal forests allows incorporating seasonal variability of land­cover classes during the rainy and dry seasons, as well as transitions between seasons. The most important variables that contributed to the accuracy were the Red, Near Infrared (NIR) and Short­Wave Infrared (SWIR) bands in single­date multispectral classification, and the months in the dry season were the most relevant in multi­date NDVI classification.