Supervised Machine Learning Classification of Urban Forest Physiognomy: Contribution of Geomorphometric and Spectral Attributes

This study aims to evaluate the relative contribution of spectral and geomorphometric attributes in the supervised classification of forest physiognomies in urban remnants of the Atlantic Forest, using machine learning algorithms. The performance of three algorithms—Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM)—was compared. The dataset consisted of twelve attributes, including spectral variables (bands and vegetation index) and non-spectral variables (slope and vegetation height), derived from very high spatial resolution data of forests located in the municipality of Niterói, part of the Rio de Janeiro Metropolitan Region, which encompasses approximately 50% of its territory. Three classification tests were conducted, and the Random Forest (RF) algorithm achieved the best performance, with an overall accuracy of 96% and F-score values above 90%. In comparison, the Support Vector Machine (SVM) obtained an overall accuracy of 88% and F-score values exceeding 74%. Non-spectral attributes, particularly elevation and vegetation height, showed the greatest importance in classification, with permutation indices of 0.48 and 0.12, respectively, while the remaining attributes scored below 0.04. The supervised classification distinguished three physiognomic classes: tall forest, medium forest, and low forest. Beyond demonstrating the superior performance of RF in classifying forest remnants, this study highlights the relevance of geomorphometric variables—especially elevation—in characterizing vegetation physiognomy, whether due to anthropogenic influence or the role of topography in plant physiology.