A Comparative Study of Training Modifications for Small Object Detection in Satellite-Based Search and Rescue Missions

Small object detection in high resolution satellite imagery for search and rescue (SAR) operations remains chal- lenging, with targets sometimes 3-4 pixels in width, compared to full images of 1000-pixel resolution. Using the SaRNet dataset containing 2,552 satellite images from a real missing person search, we evaluated three modifications to a baseline Faster R-CNN Feature Pyramid Network architecture to improve the recall performance metric on small object detection. We tested (A) Focal Loss integration to address class imbalance since targets represent <0.16% of image area, (B) multi-scale training and testing at higher image resolutions (10-20% up-scaled) and (C) decreased anchor sizes. Results were mixed. Focal Loss was the only successful modification, improving small object recall by 4.4 percentage points (10.4% relative improvement) while also increasing recall on large objects. Surprisingly, both anchor optimization and multi-scale training degraded performance despite theoretical justification. Optimized anchor sizes decreased recall across all object sizes and caused the worst AR-d20 per- formance drop (-12.64 points), revealing that geometric anchor coverage doesn’t guarantee detection improvement in transfer learning contexts. Multi-scale training decreased medium- sized object recall by 9.5 percentage points, contradicting recent super- resolution research. This work provides the first systematic evaluation of modifications of the baseline model for the SaRNet dataset towards improved small object detection. For operational SAR systems where lives depend on detection performance, our results recommend Focal Loss integration while cautioning against modifications that disrupt pre-trained model configura- tions.