A combined approach to pixel-wise classification of satellite images based on LBP, pseudocolor features, and XGBOOST

Abstract

The subject of the article is pixel-wise classification of Sentinel-2 satellite imagery represented as three-channel data mapped to the RGB color space for convenient visualization, with specific attention to the challenges posed by sensor noise and lossy compression artifacts typical for satellite data. The goal is to develop and validate a classification approach that maintains high accuracy under substantial noise and compression, by combining Local Binary Patterns (LBP) texture descriptors with pseudocolor features and employing an efficient ensemble classifier. The tasks to be addressed are: to design a compact feature representation that integrates LBP-based texture information with pseudocolor; to train and tune an XGBoost classifier on these features and compare its performance with baselines that rely on pseudocolor information alone and with simple neural network models; to assess robustness to noise and compression artifacts across a range of compression levels. The methods used include extraction of LBP descriptors to capture local texture patterns, construction of pseudocolor features from RGB-mapped Sentinel-2 channels, and concatenation of these descriptors into joint feature vectors. An XGBoost algorithm is employed to build the classification model. Model effectiveness is evaluated using the F1 score as the primary metric under varying noise and compression conditions. Visual inspection of the resulting classification maps is used to corroborate quantitative results and to analyze spatial consistency and error patterns. Conclusions. The scientific novelty of the results is as follows: for the first time in the context of Sentinel-2 pixel-wise classification, the use of LBP in combination with XGBoost has been systematically investigated and substantiated for BPG lossy compression scenarios at the optimal operating point (OOP) or nearby; it has been experimentally established that there is a substantial gain in classification accuracy for heterogeneous classes (urban areas, vegetation, bare soil) and a limited gain for homogeneous ones (water), and interaction artifacts of BPG+LBP on homogeneous surfaces have been documented, with directions outlined for adapting LBP parameters to mitigate them; the computational suitability of the approach (feature extraction, training, and classification time) for operational pipelines has been demonstrated; a comparison with a simple neural network has been conducted showing higher stability of the proposed approach on texture-rich classes under noise and compression, thereby delineating the limits of applicability of alternative methods. The study also shows that accounting for compression effects is important for operational processing pipelines: compressing images to an optimal operating point can reduce data volume and, in some cases, slightly improve classification accuracy by attenuating noise.