Deep Learning-based Fishing Activity Area Prediction with VIIRS Boat Detection Data and an Adaptive Multi-Temporal Integration Module (AMTIM)

Abstract

Deep learning-based methods for the daily prediction of fishing vessel activity areas in the Japan Sea and East China Sea (25–40° N, 120–140° E) were investigated herein, using data obtained from Visible Infrared Imaging Radiometer Suite (VIIRS) boat detection (VBD). VBD grid images were created by aggregating fishing vessels’ position data into 64×64 grids, with the final input day’s distribution used as the baseline for evaluation. To suppress detection failures from clouds, the evaluation was limited to clear detections using quality flags (QFs). Indices were binarized using thresholds of 1, 5, and ≥10 fishing vessels, and F1 and intersection over union (IoU) metrics were applied to evaluate the models’ prediction performance for congested areas. Performance improvements obtained via encoder modifications were confirmed, with U-Net models using six-channel input DenseNet121 and ResNet50 showing high performance. The 6ch_densenet121 model achieved approx. 30% improvement in the average F1 and IoU values compared to the baseline, with 5%–10 % improvements for higher vessel-density areas. A novel Adaptive Multi-Temporal Integration Module (AMTIM) was developed to integrate temporal features including the year, month, day, lunar age, solar longitude, and tidal phase. In a six-channel ResNet50 model, the use of the AMTIM improved the average F1 from 0.383 to 0.393 and the average IoU from 0.237 to 0.245, demonstrating an approx. 3 % performance gain. When the AMTIM used only basic temporal information (i.e., year, month, day), the performance degraded, confirming the importance of domain-specific temporal features. Although it was developed for the prediction of fishing activity, the AMTIM’s general-purpose design enables applications to other domains that require non-spatial temporal information integration. The models successfully complemented the activity areas during cloud-interrupted observations, confirming their contribution to continuous fishing vessel monitoring.