https://deeplearning.earth/tags/sar/Recent content in Sar on DeepLearning.EarthHugo -- gohugo.ioenThu, 01 Oct 2026 06:00:00 +0700https://deeplearning.earth/posts/2026-10-01_hrsid_sar_ship_benchmark/Thu, 01 Oct 2026 06:00:00 +0700https://deeplearning.earth/posts/2026-10-01_hrsid_sar_ship_benchmark/Two SAR ship benchmarks, two different jobs. SSDD is the small-chip sanity check (~90% held-out). HRSID (Wei et al., IEEE Access 2020) is the larger, higher-resolution set: 5,604 800×800 chips, 16,951 ships, Sentinel-1B / TerraSAR-X / TanDEM-X, 0.5–3 m. oriented-det v0.3 loads it natively (dataset.format: hrsid). A full Faster R-CNN 1× finetune from the DOTA Hub reached 78.55% mAP50 on official held-out test — rotated IoU on min-area rectangles. There is no HRSID Hub slug.https://deeplearning.earth/posts/2026-09-28_ssdd_sar_ship_finetune/Mon, 28 Sep 2026 06:00:00 +0700https://deeplearning.earth/posts/2026-09-28_ssdd_sar_ship_finetune/Optical pretrain, SAR finetune, twelve epochs. SSDD (Zhang et al.) is 1,160 SAR ship chips (~190–688 px), RadarSat-2 / TerraSAR-X / Sentinel-1, 1–15 m. Official split: file numbers whose last digit is 1 or 9 are test (~232); the rest are train (~928). oriented-det v0.3 loads it natively (dataset.format: ssdd). A full Faster R-CNN 1× finetune from the DOTA Hub reached 90.34% mAP50 on held-out test. There is no SSDD Hub zoo.