INFORMATION TECHNOLOGY FOR AUTOMATED ANALYSIS OF UAV AERIAL IMAGES BASED ON SPLINE APPROXIMATION OF TEXTURE FEATURE DISTRIBUTION

DOI: 10.31673/2412-4338.2026.019022

Abstract

The article presents an information technology for automated analysis of unmanned aerial vehicle (UAV) imagery based on polynomial spline approximation S2,0. The proposed solution overcomes the high computational complexity of modern neural network approaches and the inadequacy of parametric models in describing heterogeneous textures of real aerial images. The developed information technology is implemented as a three-level modular Core–Services–GUI architecture with a two-phase computational pipeline whose theoretical complexity is O(Aμ · Aσ), regardless of the input image resolution. The first phase approximates the distribution of texture features μ × σ using a polynomial spline on a regular grid; the second phase performs a topographic search for modes, verifies them using the analytical Hessian H(S), and selects the most reliable pseudo-labels according to the parameter pauto. The scientific novelty lies in the formalization of two deterministic clustering strategies: geometric partitioning of the feature space (Discr) through TLS approximation of discriminant lines and the use of convergent potential (Grad). Guaranteed determinism of the results (σ(ARI) = 0) ensures stable reproducibility of the training dataset across different labeling sessions. Empirical verification was performed on 120 realizations of the discrete intensity field from three aerial image datasets (BPLA, LoveDA, OpenEarthMap). It was found that at pauto = 0.10–0.20, the quality of the selected pseudo-labels reached ARI = 0.571–0.580, which is twice as high as that of competing methods without a confidence mechanism, while selecting the 50% most confident pixels outperformed oracle methods with a priori knowledge of the number of classes across all three datasets. The accuracy of automatic cluster number estimation, |Δk| = 0.942–0.950, is three times higher than that of GMM_BIC. The log-log slope α = 0.070 confirms the suitability of the technology for SWaP-constrained onboard deployment, while the determinism σ(ARI) = 0 guarantees reproducibility of the training dataset across labeling sessions. The statistical significance of the results was confirmed using the Friedman test (p < 0.05) and Cohen's effect size (d > 0.8). The proposed information technology provides effective anomaly detection (ROC-AUC = 0.630) and represents a comprehensive, balanced solution for automated aerial surveillance systems.

Keywords: UAV, aerial imagery, spline approximation, texture features, clustering, anomaly detection, semantic segmentation, machine learning, unsupervised learning.