Obtaining agricultural land cover in Sentinel-2 satellite images with drone image injection using Random Forest in Google Earth Engine

To obtain accurate information on land cover changes in the agricultural sector, we propose a supervised classification method that integrates Sentinel-2 satellite imagery with images surveyed from Remote Piloted Aircraft Systems (RPAS). The methodology was implemented on the Google Earth Engine platform. Initially, the Sentinel-2 imagery collection was integrated into a single image through a median reduction process. Subsequently, the high-pass filter (HPF) pansharpening image fusion method was applied to the thermal spectral bands to obtain a final spatial resolution of 10 m. To perform the integration of the two image sources, the RPAS image was normalized by using a 5X5 gaussian texture filter and the pixel was resampled to five times its original size. This procedure was performed iteratively until reaching the spatial resolution of the Sentinel-2 imagery. Besides, the following inputs were added to the classification: the spectral indices, calculated from the Sentinel-2 and RPAS bands (e.g. NDVI, NDWI, SIPI, GARI), altimetric information and slopes of the zone derived from the SRTM DEM. The supervised classification was done by using the Random Forest technique (Machine Learning). The land cover seed reference to perform the classification was manually captured by a thematic expert, then, this reference was distributed in 70% for the training of the Random Forest algorithm and in 30% to validate the classification. The results show that the incorporation of the RPAS image improves thematic accuracy indicators by an average of 3% compared to a classification made exclusively with Sentinel-2 imagery.