Repository navigation
Home
CAD-grade photovoltaic array segmentation, 3D roof plane fitting, physical BIM module packing, annual solar yield forecasting, and 4D temporal change tracking from aerial orthophotos. Developed and maintained by Valeford under the Spacement brand.
Spacement PVExtent is released in two production tiers:
-
PVExtent Vidda (v1.0 Standard):
- DeepLabV3+ ResNet50 ASPP with Photometric Invariance
- Fast single-pass sliding-window inference (~15 ms/tile)
- Calibrated posterior probabilities (
$T^* = 0.4068, ECE = 0.47%$ ) - Ideal for nationwide / regional municipal screening and edge/laptop execution
-
PVExtent Tinden (v1.0 High-End Flagship):
- Heterogeneous Production Ensemble (Photometric V5 + Self-Trained NRW V6)
- 8-Fold Dihedral Symmetry Group (
$D_4$ ) TTA + Multi-Scale Pyramid ($0.9\times, 1.0\times, 1.1\times$ ) - Seamless 2D Separable Hann Apodization (
$W = \sin^2 \otimes \sin^2$ ) - Bayesian Epistemic & Aleatoric Uncertainty Decomposition
- 89.15% Test IoU, 94.26% Dice, 96.32% Recall, 99.955% Specificity across sealed 421 benchmark crops
- 3D Digital Surface Model (DSM) Plane Fitting & PVGIS Yield Transposition
- 4D Multi-Temporal Change Detection (Zubau, Erweiterung, Rückbau, Bestand)
- 4-Band Multispectral NIR / CIR Ingestion & Green Roof (Dachbegrünung) Veto
Evaluated across 43,110,400 test pixels from the 421 strictly reserved test crops that remained sealed until final freeze:
| Model Architecture / Release | Test IoU | Test Dice (F1) | Pixel Precision | Pixel Recall | Test Pixel Accuracy | False Negatives (Missed PV) | False Positives |
|---|---|---|---|---|---|---|---|
| Baseline V1 (ResNet18 U-Net) | 85.92% | 92.42% | 93.98% | 90.92% | 97.08% | 767,417 px | 492,078 px |
| PVExtent Vidda (V5 Photometric) | 88.16% | 93.71% | 93.89% | 93.53% | 97.54% | 546,934 px | 514,455 px |
| DeepLabV3+ V6 (NRW Self-Trained) | 89.07% | 94.22% | 92.21% | 96.32% | 97.68% | 311,121 px | 687,495 px |
| PVExtent Tinden (Ensemble + |
89.15% | 94.26% | 93.61% | 94.93% | 97.74% | 428,375 px | 548,039 px |
- PVExtent Vidda vs. Tinden Guide: Detailed feature comparison and deployment specs.
- 3D Elevation & Solar Yield Audit: True 3D surface area expansion ($A / \cos(\theta)$), roof pitch/azimuth fitting, PVGIS yield transposition.
-
Multi-Temporal 4D Change Detection Audit: Bipartite matching across survey flights (Zubau, Erweiterung, Rückbau, Bestand) with net capacity accounting (
$\Delta\text{kWp}, \Delta\text{Module}, \Delta\text{m}^2$ ). - Multispectral NIR & Spectral Signature Audit: 4-band RGB+NIR radiometry, NDVI/PVSI biophysical indices, CIR false-color rendering, green roof false-positive suppression.
- Sealed 421 Test Benchmark Audit: Formal peer-level verification on untouched test crops.
- CAD Vectorization & Manhattan Regularization Audit: 90° rectilinear boundary regularization and physical 400W BIM module layout.
WICHTIGER HINWEIS: Alle Modelle (
PVExtent Vidda,PVExtent Tinden), Berechnungen, Vektorgeometrien und Ertragsprognosen werden ausdrücklich "AS IS" OHNE JEGLICHE GEWÄHRLEISTUNG bereitgestellt. Valeford und Spacement schließen jegliche Haftung für wirtschaftliche, bauliche, statische, behördliche oder rechtliche Folgen vollständig aus.
Developed and maintained by Valeford under the Spacement brand. Original project code: Apache-2.0. See the repository NOTICE for attribution.
Spacement PVExtent
- Home
- PVExtent Vidda vs. Tinden (v1.0 Releases)
- Getting started
- Dataset format
- Annotation guide
- PV4GER-SEG import
- Import audit
- Geographic splits
- Split audit
- Label review
- NRW candidates
- Bulk NRW training imagery
- Bulk mask curation
- Live training monitor
- Training quality exclusions
- Final bulk mask audit
- Experimental negative curation
- Negative curation audit
- Exploratory negative holdout
- Negative holdout audit
- Exploratory positive holdout
- Positive holdout audit
- Training smoke test
- Smoke audit
- Pretrained baseline training
- Baseline comparison
- Baseline comparison audit
- Six-epoch continuation audit
- V4 bulk training audit
- Domain-balanced training audit
- DeepLabV3+ ASPP architecture audit
- DeepLabV3+ Focal + Lovász loss audit
- DeepLabV3+ V5 Photometric Invariance audit
- CAD vectorization & rectilinear audit
- Test-Time Augmentation (TTA) audit
- Probability calibration & temperature scaling audit
- Seamless overlapping tiled inference & Hann blending audit
- Heterogeneous ensemble & quality triage audit
- Sealed 421 test benchmark audit
- 3D elevation & solar yield audit
- Multi-temporal 4D change detection audit
- Multispectral NIR & spectral signature audit
- Encoder weights and licenses
- Baseline audit
- Frozen validation errors
- Validation error audit
- Candidate predictions
- Research roadmap
Project