Influence of local binary pattern configurations and XGBoost on the quality of satellite image classification with noise and compression: experimental study

Abstract

The paper addresses the relevant task of pixel-wise classification of multispectral satellite images under conditions of critical data quality degradation caused by additive sensor noise and lossy compression artifacts. A method for spatial feature extraction based on Local Binary Patterns (LBP) using various configurations (1:8, 2:16, 3:24) and their combinations to improve class separation accuracy is considered. An experimental study was conducted on Sentinel-2 image fragments (Kharkiv region) distorted by additive Gaussian noise with a variance of σ² = 100 and subjected to per-channel compression using the BPG encoder. The XGBoost decision tree ensemble was used as a classifier. Simulation results confirmed that the application of mono-scale LBP patterns is insufficient for reliable segmentation of heterogeneous objects under conditions of strong noise. It is proven that forming an extended feature vector by concatenating multi-scale LBP configurations (1:8, 2:16, 3:24) ensures an increase in the F1-score metric to 0.9530, which exceeds the indicators of basic configurations by more than 1%. A detailed analysis of metric dynamics by class revealed that the "Water" class demonstrates the highest stability (F1 > 0.99) due to spectral homogeneity. At the same time, for structurally complex classes "Urbanization" and "Vegetation," incorporating large-radius features proved critically important, allowing to minimize the influence of local brightness fluctuations and stabilize the AUC metric at a level > 0.99. An important empirical result was the detection of a positive effect of compression on the classification accuracy of noisy images: coefficient quantization by the BPG encoder acted as a low-pass filter, partially compensating for the high-frequency component of Gaussian noise. Additional research around the optimal operating point revealed that the method remains robust under moderate changes (Q=31). Furthermore, it continues to exhibit high stability and maintain segmentation accuracy even with a substantial amplification of compression artifacts (Q=43), thereby confirming the reliability and efficiency of the algorithm throughout the entire evaluated range of compression distortions. It was established that the integration of texture features of different scales allows the XGBoost model to form robust decision rules, while the optimal balance between computational complexity and accuracy is achieved when limiting ensemble parameters (n_estimators=200, max_depth=8), as further model complexity does not lead to a statistically significant increase in recognition quality.

Author Contributions: Conceptualization – M. A. Rybnytskyi and S. S. Kryvenko; Methodology – M. A. Rybnytskyi; Formal Analysis and Data Processing – M. A. Rybnytskyi; Investigation – M. A. Rybnytskyi; Writing – Original Draft Preparation – M. A. Rybnytskyi; Writing – Review & Editing – M. A. Rybnytskyi and S. S. Kryvenko. All authors have read and agreed to the published version of the manuscript.

Funding: This research received no external funding.

Data Availability Statement: Data available on reasonable request from the authors.

Acknowledgments: The authors express sincere gratitude to the reviewers and editors for their valuable comments, recommendations, and attention to the work.

Conflicts of Interest: The authors declare no conflict of interest.