УДОСКОНАЛЕННЯ МЕТОДУ ПОПІКСЕЛЬНОЇ КЛАСИФІКАЦІЇ ДАНИХ ДИСТАНЦІЙНОГО ЗОНДУВАННЯ В УМОВАХ СПОТВОРЕНЬ НА ОСНОВІ КОМБІНОВАНИХ LBP-ОЗНАК

<jats:p>The subject of this article is the process of pixel-wise classification of satellite images under distortions caused by lossy data compression and additive white Gaussian noise (AWGN). The relevance of the work is driven by the fact that while previous studies using the Local Binary Patterns (LBP) method have demonstrated promising results, they have also revealed the high sensitivity of baseline configurations with a small feature extraction radius to additive pixel noise and compression artifacts (such as boundary blurring and the formation of block effects). In certain cases, this leads to the misclassification of homogeneous areas distorted by compression (e.g., misidentifying water surfaces as vegetation). The goal is to improve the accuracy of pixel-wise classification of satellite images under distortions by developing a robust approach based on a combined feature space that integrates spectral intensity and texture information of different scales. The tasks addressed are: conducting an experimental analysis on 4 diverse satellite scenes modeled under 4 distortion states (reference image, noisy, and two levels of compression); forming an expanded feature space using the combined LBP (by integrating radii of 1, 2, and 3 pixels); and tuning and optimizing the hyperparameters of the XGBoost ensemble classifier. The methods used are based on texture analysis tools, decision-tree-based machine learning algorithms, and statistical accuracy assessment techniques. The study's results experimentally confirm that combining LBP descriptors across different radii partially mitigates microtexture loss caused by compression. Conclusions. The findings confirm that the proposed algorithm effectively compensates for the shortcomings of individual LBP texture descriptors under image distortion conditions. The practical significance of the obtained results lies in establishing a reliable baseline for developing robust automated monitoring and analysis systems for Earth remote sensing (RS) data.</jats:p>