Impact of segmentation algorithms on multisensor LULC classification in a semiarid Mediterranean area

Abstract

Geographic object-based image analysis methods usually provide better results than pixel-based methods for classifying land use and land cover from high and medium resolution satellite imagery. This study compares the results of Random Forest (RF) and Multilayer Perceptron (MLP) when used to classify the segments obtained on an RGB+NIR Sentinel-2 image using three different segmentation algorithms, Multiresolution (MR), Region Growing (RG), and Mean-Shift (MS). The hyperparameters of these algorithms were optimised minimising the intra-object heterogeneity and maximizing the inter-object heterogeneity, integrating them in an optimization loop. Geometric and two different centrality and dispersion statistics were computed from some Sentinel-1, Sentinel-2 and LiDAR variables over the segments, and used as features to classify the datasets. The highest segment cross-validation accuracies were obtained with RF using MR segments: 0.9048 (k=0.8905), while the highest accuracies calculated with test pixels were obtained with MLP using MR segments: 0.9447 (k=0.9303), both with the mean and standard deviation of the feature set. Although the overall accuracy is quite high, there are problems with some classes in the confusion matrix and, significant misclassification appear when a qualitative analysis of the final maps is performed, indicating that the accuracy metrics may be overestimated and that a qualitative analysis of the results may also be necessary.