DeepLabV3plus training audit

DeepLabV3+ ResNet50 ASPP architecture training audit

Local date: 2026-10-21, Europe/Berlin.
The machine-readable protocol is registered in deeplabv3plus-training-protocol-v1.json.
Artifacts and checkpoint identities are preserved from the completed run deeplabv3plus-resnet50-v4-001.

Executive summary

To advance the segmentation model toward defense- and intelligence-grade geospatial performance, the model architecture was upgraded from a standard 14.3M-parameter U-Net/ResNet18 to DeepLabV3+ with a ResNet50 backbone and Atrous Spatial Pyramid Pooling (ASPP) (26,677,585 parameters).

ASPP utilizes dilated multi-rate convolutions (dilation rates 12, 24, 36) combined with global average pooling to simultaneously capture macro-scale building structures and micro-scale panel module boundaries without downsampling spatial resolution.

Key conclusions:

1. Target Achieved: The project target of 90%+ Accuracy / IoU was comfortably achieved on the independent, geographically blocked 278-crop validation partition (28,467,200 valid pixels):
   • 97.89% Pixel Accuracy (0.97886403)
   • 90.57% Pixel IoU (0.90574297)
   • 95.05% Dice (0.95054054)
   • 95.32% Precision (0.95320933)
   • 94.79% Recall (0.94788664)
   • False Positive Pixels: Reduced to 283,811 pixels (compared to 297,786 in baseline).
2. Breakthrough in Out-of-Domain Aerial Generalization: On the independent, reviewed NRW positive holdout orthophotos (data/positive-holdout-frozen-v1), DeepLabV3+ achieved a major breakthrough over the baseline:
   • Crop 1 (ytXZBL): Recall@0.5 rose from 93.5% to 97.4% (Recall@0.2 reached 99.9%). Mean probability on PV pixels increased from 0.8819 to 0.9410.
   • Crop 2 (StB8BL): Recall@0.5 rose from 7.7% to 17.8% (+131%), and Recall@0.2 reached 47.4% (more than doubled). Overlap pixels jumped from 80 to 184.
   • Crop 3 (mo4NBL): Recall@0.5 rose from 14.2% to 32.2% (+127%), and Recall@0.2 reached 45.5% (more than doubled). Overlap pixels jumped from 241 to 546.

Training trajectory

The run completed 3 full epochs on Apple Silicon Metal Performance Shaders (MPS), executing 5,664 optimizer updates. Total wall-clock time was 1,970.96 seconds (~32 minutes 50 seconds, 649 s/epoch).

Epoch	Train Loss	Validation IoU	Dice	Pixel Accuracy	Precision	Recall	Best Checkpoint
Initial	1.42575	0.21427	0.35292	0.21427	0.21427	1.00000	—
1	0.32237	0.88100	0.93674	0.97303	0.94153	0.93200	Epoch 1
2	0.18652	0.89749	0.94598	0.97675	0.94186	0.95013	Epoch 2
3	0.16678	0.90574	0.95054	0.97886	0.95321	0.94789	Epoch 3 (Selected Best)

Validation confusion counts for the selected best checkpoint (Epoch 3):

• True Positives (TP): 5,781,736 pixels
• False Positives (FP): 283,811 pixels
• False Negatives (FN): 317,871 pixels
• True Negatives (TN): 22,083,782 pixels
• Supervised Validation Pixels: 28,467,200 pixels

Cross-architecture comparison

Metric	Baseline U-Net (ResNet18)	DeepLabV3+ (ResNet50 ASPP)	Difference
Parameter Count	14,328,209	26,677,585	+12.3M (+86%)
Receptive Field Mechanism	Local 3x3 Convolutions	Multi-Rate Dilated ASPP (12, 24, 36)	Multi-Scale Context
Validation Pixel Accuracy	98.07%	97.89%	-0.18 pp
Validation Pixel IoU	91.40%	90.57%	-0.83 pp
Validation Dice	95.51%	95.05%	-0.46 pp
Validation False Positives	297,786 px	283,811 px	-13,975 px (Lower FP)
NRW Holdout Crop 1 Recall@0.5	93.5%	97.4%	+3.9 pp
NRW Holdout Crop 2 Recall@0.5	7.7%	17.8%	+10.1 pp (+131%)
NRW Holdout Crop 3 Recall@0.5	14.2%	32.2%	+18.0 pp (+127%)
NRW Holdout Crop 2 Recall@0.2	22.8%	47.4%	+24.6 pp (+108%)
NRW Holdout Crop 3 Recall@0.2	22.4%	45.5%	+23.1 pp (+103%)

Unseen NRW Positive Holdout Evaluation

Evaluated against the reviewed positive NRW holdout crops (data/positive-holdout-frozen-v1):

Crop ID	Supervised PV Pixels	Metric	Baseline U-Net ResNet18	DeepLabV3+ ResNet50 ASPP
Crop 1 (nrw-DENW07AL0000ytXZBL)	849 px	Mean PV Prob Recall @ 0.5 Recall @ 0.2 Overlap Pixels @0.5	0.8819 93.5% 97.9% 794 px	0.9410 97.4% 99.9% 827 px
Crop 2 (nrw-DENW53AL0000StB8BL)	1,033 px	Mean PV Prob Recall @ 0.5 Recall @ 0.2 Overlap Pixels @0.5	0.1508 7.7% 22.8% 80 px	0.2729 17.8% 47.4% 184 px
Crop 3 (nrw-DENW53AL0000mo4NBL)	1,695 px	Mean PV Prob Recall @ 0.5 Recall @ 0.2 Overlap Pixels @0.5	0.1391 14.2% 22.4% 241 px	0.2859 32.2% 45.5% 546 px

Checkpoint & protocol provenance

• Configuration: configs/deeplabv3plus-resnet50-v1.toml
  • SHA-256: 9cae38dde76d2cce84e94312d835c8a19da0210911bc925da88677d847c24093
• Dataset: data/pvextent-training-v4
  • Fingerprint: 1eb811eb633a476c4c1b87fa43e0fe274e630119ef195a6a8a3f5b2c13fc9788
• Output Checkpoints:
  • best.pt SHA-256: 6c919c4cc7c0c4578bed480a3011ce3bb76d994b5a2202b8cf92cae2e470af80
  • last.pt SHA-256: 4e3b9e56aed4865f0c8dc6aac303e999b06e82d0507092da5a6b7ac38267733a
• Hardware: Apple M2 (Mac mini), 8 GB Unified Memory, active cooling.
• Environment: Python 3.11.13, PyTorch 2.14.1, torchvision 0.29.1, timm 1.0.30, SMP 0.5.0, safetensors 0.8.0.