Remotely sensed imagery and machine learning for mapping of sesame crop in the Brazilian Midwest

The changes in landscapes have been followed more intensely in recent decades thanks to scientific advances, both in the field of technological improvement of satellites and in remote sensing techniques. Advanced and efficient machine learning techniques have helped remote sensing professionals to determine these changes, from the simplest to the most complex landscapes, allowing the identification of the most varied land uses and occupation, as well as the estimation of the areas that these uses occupy, allowing a more dynamic management of natural resources, especially in agricultural exploitation, providing reliable information to decision makers. Thus, the objective of this work is, through machine learning techniques, to estimate the area of sesame (Sesamum indicum) cultivation in the crop season 2021/2022, in the municipality of Canarana, in the state of Mato Grosso, comparing the performance of the Random Forest and Support Vector Machine (SVM) classifiers, using images from the Landsat 8/OLI satellite. As a source of information for the supervised classification, control points in geographic coordinates were collected in the study area to identify the areas cultivated with sesame. The vegetation indices NDVI, EVI, NDBI, PVI and SAVI were used for the elaboration of thematic maps, along with the Landsat 8/OLI images. Global Accuracy and Kappa index were used as a rule of thumb in the evaluation of the thematic maps, compared by the Z test, with significance at & alpha; = 0.05. The test revealed that the Random Forest classifier showed better performance in identifying the sesame cultivated areas, with Global Accuracy of 0.95 and Kappa of 0.90, when compared to SVM, which showed 0.91 and 0.81, respectively. The use of machine learning techniques in Landsat 8/OLI images proved satisfactory in estimating areas cultivated with sesame in the municipality of Canarana-MT, demonstrating confidence in the mapping. The way Random Forest structures its training model, creating as many decision trees as necessary, ended up mitigating more classification errors , proved to be more promising when compared to SVM. As a rule, both algorithms showed potential for mapping the sesame crop.