Instance-level scene detection on MethaneSAT cross-validation (test)
22TP (True Positives)Mask R-CNN
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Mask R-CNNStrategy=Fine-Tuning, Mode=Baseline, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 22 | 15 | 0 | 92 | 74 | 60 | 98 | |
| Mask R-CNNStrategy=Joint, Mode=Baseline, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 22 | 23 | 0 | 95 | 65 | 49 | 98 | |
| Mask R-CNNStrategy=Curriculum, Mode=Baseline, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 22 | 24 | 0 | 93 | 65 | 49 | 98 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Fine-Tuning, Mode=High Sensitivity, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 21 | 8 | 1 | 91 | 81 | 71 | 94 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Joint, Mode=High Sensitivity, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 21 | 15 | 1 | 93 | 72 | 58 | 94 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Curriculum, Mode=High Sensitivity, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 21 | 13 | 1 | 92 | 74 | 62 | 94 | |
| Mask R-CNNStrategy=MethaneSAT-only, Mode=Baseline, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=Trained from scratch2026.05 | 18 | 63 | 4 | 73 | 35 | 23 | 80 | |
| Mask R-CNN with Physics-informed post-processingStrategy=MethaneSAT-only, Mode=High Sensitivity, Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=Trained from scratch2026.05 | 16 | 47 | 6 | 67 | 38 | 28 | 71 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Joint, Mode=High Precision (QND), Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 16 | 5 | 6 | 71 | 76 | 78 | 74 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Curriculum, Mode=High Precision (QND), Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 16 | 4 | 6 | 70 | 76 | 81 | 73 | |
| Mask R-CNN with Physics-informed post-processingStrategy=Fine-Tuning, Mode=High Precision (QND), Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=MethaneAIR pre-trained2026.05 | 15 | 1 | 7 | 68 | 79 | 92 | 70 | |
| Mask R-CNN with Physics-informed post-processingStrategy=MethaneSAT-only, Mode=High Precision (QND), Backbone=ResNet-50, IoU threshold=0.1, Patch size=768×768, Pre-training=Trained from scratch2026.05 | 12 | 9 | 10 | 58 | 56 | 61 | 56 |