Аңдатпа
Google Earth Engine (GEE) is widely used in Land Use and Land Cover (LULC) mapping studies due to its cloud-computing infrastructure and broad access to geospatial datasets. The platform provides several supervised classification algorithms and enables the implementation of customized methods through its programming interface. Among the classifiers commonly used within the GEE environment, Random Forest (RF) stands out for its robustness, ability to handle heterogeneous data, and strong performance in supervised classification applications. Despite the consolidation of this methodology, different input variables can be incorporated into the classification process, potentially influencing the accuracy of results. In this context, the present study evaluates the influence of auxiliary variables on the accuracy of land use and land cover mapping in the São Francisco Verdadeiro River Basin, located in western Paraná, Brazil. Two distinct classification approaches were tested: the first method relied exclusively on the spectral bands of the MSI sensor onboard the Sentinel-2 satellite (B1, B2, B3, B4, B5, B7, B8, B8A, B9, B11, and B12); the second method incorporated, in addition to the spectral bands, spectral indices(SR, NDVI, NDWI, SAVI, and EVI), Spectral Mixture Analysis, and topographic variables (slope and elevation). The processing used two Sentinel-2 Level-2A scenes (surface reflectance atmospheric correction) acquired on October 17, 2022, and was carried out on the GEE platform using the RF classifier. The results demonstrated the superiority of the second approach, which achieved 90.4% overall accuracy based on an error matrix assessment using field data. In the comparison between methods, the upper section of the watershed exhibited the greatest spatial discrepancies, whereby, while the approach based solely on spectral bands overestimated natural vegetation areas, the inclusion of auxiliary variables in Method 2 enabled better detection of straw/crop residue areas. The variable importance analysis confirmed that spectral indices contributed to improved classification performance in the second method.