This study focuses on determining dynamic threshold coefficients for crater detection on war-damaged fields in Ukraine using regression models and Sentinel-2 satellite images. Based on expert-labeled data for threshold coefficients across 360 fields in Donetsk region, models based on Multiple Linear Regression, Random Forest, and Support Vector Machine were trained and compared for accuracy. The input parameters included the publication date of the satellite image, land cover class and statistical indicators (minimum, average, maximum value of spectral bands B2 (blue), B3 (green) and vegetation indices NDVI and GCI). The best predictor combination for each model was selected using sequential feature selection with 3-fold cross-validation. The results identified Random Forest regression as the most accurate model for determining thresholds for each band and index. R \(^2\) ranged from 0.72 (threshold for GCI(1)) to 0.96 (threshold for GCI), with corresponding RMSE of 0.171 and 0.039, respectively. Using predicted threshold coefficients from the models resulted in an average reduction of false positives rate by 8% while simultaneously decreasing true positives rate by 13% during crater detection. Future efforts will involve expanding and improving the quality of the training dataset and using more sophisticated machine learning models to improve the results.