DATA FUSION FOR LOCAL CLIMATE ZONES CLASSIFICATION BASED ON ENSEMBLE OF CLASSIFIERS

Multisource remote sensing data provide information of high relevance for classification and climate studies in urban areas and are of particular interest for regional and global climate science. To classify the urban environment using predefined Local Climate Zones we show a methodology that uses feature extraction from the multisource data and image segmentation as input data to an ensemble of classifiers and verify which algorithm has the best accuracy. The algorithms used were AdaBoost, Random Forest, Multi-layer Perceptron and an ensemble of those classification methods. The multispectral images were from Landsat 8 and Sentinel 2 resampled to 100m. LCZs were generated for Paris and Sao Paulo and the visual analysis and quantitative testing of results show the ensemble of classifiers had the best result for both cities with OA of 87.7% and 83%, Kappa of 0.81 and 0.80, for Paris and Sao Paulo, respectively.