Bdappv Models

Models

Two architectures, trained independently on each imagery provider:

Note on training data and licensing. Checkpoints fine-tuned on the Google subset of BDAPPV (*_google_*) derive from imagery distributed under CC-BY-NC 4.0; commercial users should prefer the IGN-trained checkpoints (*_ign_*, CC-BY 4.0 imagery) or assess accordingly. All models are initialized from Mayer et al. (2022) checkpoints — see their repository for base model licensing.

Benchmark protocol

Three evaluation tracks are defined:

Track 1 — Segmentation (single provider) Train and evaluate on the same provider. Report IoU and F1 on the test split.

Track 2 — Classification (single provider) Train and evaluate on the same provider. Report accuracy and F1 on the test split.

Track 3 — Distribution shift (cross-provider) Train on Google, evaluate on IGN test split. This is the primary robustness benchmark. Report IoU.

Rules:

• The test split must not be used for model selection or hyperparameter tuning — validation split only.
• The spatial holdout by department must not be modified. Re-splitting invalidates comparability with published results.
• For Track 3, only the Google training split may be used for training.

Results

Models evaluated on the official test split (spatial holdout by French department — see dataset card for details).

Segmentation (DeepLabV3-ResNet101)

Classification (InceptionV3)

Note on classification cross-provider results: the IGN-trained model collapses on Google imagery (Recall=0.08, Precision=1.0), indicating the model rarely predicts positives — a degenerate operating point. This illustrates the severity of the distribution shift documented in Kasmi et al. (2025).

Usage

A model.py helper is included in this repo to simplify loading:

from huggingface_hub import hf_hub_download
import importlib.util

path = hf_hub_download("gabrielkasmi/bdappv-models", "model.py")
spec = importlib.util.spec_from_file_location("bdappv_model", path)
mod  = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

seg = mod.load_segmentation_model("google")    # or "ign"
clf = mod.load_classification_model("google")  # or "ign"

Both functions return the model in eval() mode. An optional device argument is supported ("cpu", "cuda", "mps").

Training

Models trained on the official BDAPPV splits using:

• Optimizer: AdamW (lr=1e-4, weight_decay=1e-4)
• Scheduler: Cosine annealing
• Effective batch size: 32 (batch 16 × grad accum 2)
• Early stopping: patience=7 epochs on validation metric
• Input size: 400×400 px
• Initialization: checkpoints from Mayer et al. (2022), who fine-tuned DeepLabV3-ResNet101 and InceptionV3 on 10 cm/px orthoimagery from North Rhine-Westphalia (Germany) for rooftop PV detection. These checkpoints were then further fine-tuned on BDAPPV using the splits above.

Training scripts available in the BDAPPV dataset repository.

Citation

If you use these models, please cite:

@article{kasmi2022towards,
  title={Towards unsupervised assessment with open-source data of the accuracy of deep learning-based distributed PV mapping},
  author={Kasmi, Gabriel and Dubus, Laurent and Blanc, Philippe and Saint-Drenan, Yves-Marie},
  journal={arXiv preprint arXiv:2207.07466},
  year={2022}
}

References

• Mayer et al. (2022). 3D-PV-Locator: Large-scale detection of rooftop-mounted photovoltaic systems in 3D. Applied Energy, 310, 118469. (source of the base checkpoints)
• Kasmi et al. (2023). A crowdsourced dataset of aerial images with annotated solar photovoltaic arrays and installation metadata. Scientific Data, 10, 59. (BDAPPV dataset)
• Kasmi et al. (2025). Space-scale exploration of the poor reliability of deep learning models: the case of the remote sensing of rooftop photovoltaic systems. Environmental Data Science. (cross-provider distribution shift)

inception_google_best.pth

pth · 93.2 MB · SHA-256 ba031e7d0eac…b008 · Hugging Face