Detection of Three Perennial Crops with Two Supervised Classification Methods from Landsat-8 and Sentinel-2 Satellite Images

Abstract

In order for there to be adequate territorial planning, an orderly occupation of the territory is necessary; The main input is reliable information on land cover. Therefore, in recent times, new remote sensing methods have been born that need to be tested. This research compared the accuracy of the traditional remote sensing method, Maximum Likelihood (ML), versus an artificial intelligence-based method called Random Forest (RF), in the detection of banana, cocoa and palm crops (in addition to other covers). Satellite imagery from Landsat-8 and Sentinel-2 was used, which was downloaded from the USGS website; the bands in the Sentinel-2 image were atmospheric correction applied with QGIS plugins created by Congedo (2021). For the two images, the spectral indices NDVI, GNDVI, GCI and MSI were calculated, and the regions of interest (ROIs) were created on the stack of calculated indices of each image; For the creation of ROIs (both for training and validation of results) coordinates were taken in situ and from Google Earth Pro version 7.3.4.8642. ML classification was performed in ENVI 5.3 software and RF classification was performed in QGIS 3.22.7 with Dzetsaka tools created by (Karasiak, 2019). For the validation of results, the traditional confusion matrix and the Kappa coefficient proposed by Cohen (1960). The results showed that RF is slightly superior to ML, although with certain nuances. In general, the classification of Sentinel-2 with RF obtained better accuracy (89%), although for the detection of the aforementioned crops it was the least accurate.