Performance Assessment of Optical and SAR Imagery for Superpixel-Based Deforestation Mapping in the ForestEyes Project

• Vinícius D'Lucas B. e Queiroz UNIFESP
• Hugo Resende UNIFESP
• Fabio A. Faria UNIFESP / Universidade de Lisboa
• Álvaro L. Fazenda UNIFESP

Resumo

This study investigates the potential of SAR data (Cand L-band), individually and combined with optical imagery (Landsat-8 OLI), for classifying deforested areas within the ForestEyes project. Six classification scenarios with varying data combinations were defined. Images were segmented using the Simple Linear Iterative Clustering (SLIC) algorithm, and Haralick texture descriptors were extracted from each superpixel. Classifications were performed using Support Vector Machine (SVM) and Random Forest (RF), with ground truth labels derived from PRODES (Monitoring Program of the Amazon Forest by Satellite). Performance evaluation was based on 100 classification iterations, from which 95% confidence intervals were calculated for confusion matrix-derived metrics, including recall and precision. Results showed that scenarios using optical data yielded the best classification outcomes, with high recall and precision values for the Recent Deforestation class. The inclusion of SAR data did not lead to statistically significant improvements and, increased Recent Deforestation false positives, especially in scenarios relying exclusively on SAR data. Confidence interval analysis confirmed the superior performance of optical imagery for detecting deforested areas in this context.

Palavras-chave: Deforestation, Support vector machines, Image segmentation, Optical polarization, C-band, Optical imaging, Adaptive optics, Radar polarimetry, L-band, Random forests