Feature Extraction for Multispectral Analysis of Cereal Crops Using Optimized Computer Vision Pipelines 44

Abstract

The article presents the results of a study aimed at improving the stability, reproducibility, and structural consistency of computer vision pipelines for multispectral UAV imagery of winter wheat canopies. A new adaptive preprocessing model is introduced, incorporating illumination normalization based on a modified Retinex/MSRCR algorithm, entropy-regulated spatial-spectral filtering for noise suppression, and instability-driven spectral feature fusion to obtain stable multispectral descriptors. The model is formulated as a multi-objective preprocessing framework, jointly optimizing illumination invariance, noise robustness, structural preservation, and information richness. Experiments conducted on the open-access UAV dataset of nine winter-wheat fields (Switzerland) demonstrated a reduction of the coefficient of variation to 0.12 and RMSE to 0.089, together with improvements in structural similarity (SSIM = 0.923) and spectral entropy (H = 5.9), significantly outperforming classical normalization methods. The results confirm the effectiveness of the proposed approach in mitigating illumination heterogeneity and sensor-induced distortions, ensuring stable and phenologically consistent feature extraction. The developed framework can be integrated into computer-integrated and robotic precision-farming systems to enhance the reliability of automated monitoring and decision-support processes in winter-wheat production.