Desempenho de classificadores de imagens do Sentinel-2

Abstract

Population and territorial expansion processes have been a scenario of environmental degradation, and studies that are sustained in order to follow this problem.Remote sensing is an important tool for the study of these dynamics, with the classification of satellite images as an important tool for extracting information, with a strong emphasis on environmental monitoring and identification of land use and coverage features.Therefore, the Sentinel-2 satellite is a good sensor to perform works of this theme, because it has excellent spatial resolution quality, a short temporal resolution and the free availability of its data to the public.The objective of this work was to compare the performance of algorithms in the classification of images for the Sentinel-2 satellite. The study area chosen was Pratigi's APA.The selected image was acquired through the access portal of the United States Geological Survey (USGS), dated August 4, 2017, and subjected to a process of atmospheric correction by means of the dark pixel subtraction method, (MaxVer), Support Vector Machine (SVM), and Random Trees (RT) were used to compile sample segments from the nine identified classes. For performance evaluation a new set of sample segments was collected, previously identified and confirmed to be field truth through the Google Earth Pro software. Performance was measured through the analysis of the confusion matrix, global accuracy, producer and user accuracy, and Kappa and TAU indices. The analysis and comparison of the indexes used for performance evaluation indicate that the three algorithms studied presented excellent quality results, among them, the SVM classifier algorithm showing greater prominence when the classification results for each class were observed, the use of this algorithm for the classification of images of Sentinel-2.It was also observed that the Kappa and TAU indexes presented similar and excellent values for the three algorithms, recommending the use of both for the evaluation of classifying algorithms in Sentinel-2.Thus, the importance of carrying out more studies about the performance of classifiers for the Sentinel-2 satellite, may add new variables to be observed and tested for a sensor that is relatively new and not yet studied.