Параметричне налаштування автоматизованої системи виявлення пошкоджень сільськогосподарських земель на основі радарних даних

Abstract

This work is devoted to calibrating the anomaly-detection threshold in a Sentinel-1 SAR pipeline for damage detection on agricultural land. The threshold is the control parameter that defines the binary anomaly mask and therefore the feature set of the classifier. A metric sweep over the threshold reveals a trade-off: ROC AUC favours lower values, whereas the F1-score peaks at 4.5 dB. Raising the threshold from the literature default of 2.0 dB to 4.5 dB cuts the median share of false-positive anomalous pixels on undamaged fields from 68.8 % to 22.1 %, matching the physical scale of local damage.