Análise comparativa de abordagens para classificação do estádio sucessional da vegetação de um fragmento florestal da Mata Atlântica

The remote classification of the different vegetation successional stages still represents a challenging task in face of the similar spectral response of such classes. This paper is committed to evaluate the performance of both Landsat 8 and RapidEye images in the classification of successional forest stages within a patch of Mixed Ombrophilous Forest located inside the São Joaquim National Park, Santa Catarina State, south of Brazil. Three variables dataset extracted from each image were analyzed, namely; (1) one solely consisting of the spectral bands themselves; (2) a second one comprising GLCM- based texture measures derived from the spectral bands; and (3) a third one containing these two datasets and additionally two vegetation indices obtained from the Landsat-8 image and three vegetation indices from the RapidEye image. Each dataset was subject to three classifiers: random forest (RF), support vector machine (SVM), and maximum likelihood estimation (MLE or MAXVER). Results show that Kappa coefficients ranged from 0.66 to 0.88, and both user´s and producer´s accuracies were over 50%. The best result was attained with the Landsat 8 image using the third dataset and the RF classifier. Texture measures such as mean, contrast and dissimilarity were decisive for the successful classification of both images.