Improved performance of machine learning-based VIIRS night-time detection of lit fishing vessels through feature engineering

Reliable detection of lit fishing vessels from night-time satellite observations is essential not only for monitoring fishing activity but also for investigating the distribution of fishery resources when integrated with marine environmental information. However, false detections caused by anthropogenic coastal illumination remain a major challenge for vessel detection using observations from the Visible Infrared Imaging Radiometer Suite (VIIRS). In this study, a previously developed machine-learning framework for detecting lit fishing vessels from VIIRS Day/Night Band (DNB) observations was refined through feature engineering. Several new and modified features were introduced, particularly those designed to suppress land-based false positives. The improved model achieved a substantial increase in detection performance, with the area under the precision-recall curve (PR-AUC) increasing from 0.760 to 0.924. Precision improved from 0.67 to 0.786, accompanied by an approximately 12% reduction in false detections, many of which were associated with coastal illumination. SHAP analysis revealed that distance from the coastline was among the most influential predictors, ranking immediately after brightness-related features in overall importance. Independent validation using Automatic Identification System (AIS) data showed higher VIIRS-AIS matching rates while preserving the seasonal characteristics of vessel detections. This improvement is likely attributable to the reduction of coastal false positives, which decreased the total number of erroneous detections without altering the underlying seasonal signal. The improved model was subsequently applied to investigate long-term spatiotemporal patterns of lit fishing vessels in the Sea of Japan, one of the most important fishing grounds surrounding Japan. The results indicated a marked decline in low-radiance vessels after 2020, whereas high-radiance vessels tended to operate under a more selective range of surface oceanographic conditions. These findings suggest that feature-level refinements can substantially enhance the reliability of VIIRS-based vessel detection and support more robust monitoring of fishing activity, particularly in coastal and nearshore environments.