PV module detection models

PV modules detected by FT-UNetFormer + FOM (orange) on the 8 test tiles of split 0

PV modules detected by ftunetformer_fom/fold0.safetensors (orange) on the 8 test tiles of split 0, with the F1 of each tile (made with example/run_example.py --overlay of GeoSeg-PV). Imagery © 2023 Google, Airbus, CNES/Airbus, Maxar Technologies; the imagery is not covered by the CC BY licence of the weights and is not otherwise redistributed.

Trained weights for the detection of individual photovoltaic (PV) modules in very high resolution satellite imagery: 10 segmentation networks, each with and without a Fine Optimization Module (FOM), 5 weights per configuration (one per data split). They are the models of

N. Lukač, S. Seme, K. Sredenšek, G. Štumberger, D. Mongus, B. Žalik, M. Bizjak. Detection and Optimization of Photovoltaic Arrays' Tilt Angles Using Remote Sensing Data. Applied Sciences 15(7), 3598, 2025. https://doi.org/10.3390/app15073598

The code to run, evaluate and train them is GeoSeg-PV, a fork of GeoSeg.

License: Creative Commons Attribution 4.0 International (CC BY 4.0). You may use, share and adapt the weights for any purpose, also commercially, if you give credit by citing the paper (see Citation). The code that runs them, GeoSeg-PV, is under GPL-3.0.

Files

Path Content
<config>/fold<k>.safetensors Weights of configuration <config> trained on split k (0 to 4). The metadata holds arch, fom, fold, epoch, val_F1, test_F1 and test_IoU
metrics.json Per configuration and split: epoch, validation F1, test F1 and IoU, test tiles, file size
assets/overlay.png The figure above (satellite imagery © Google and partners, not CC BY)

85 files, 11.2 GB. A configuration is named after its network and encoder, with _fom if it has the FOM.

Use

git clone https://github.com/nikolukac/GeoSeg-PV.git && cd GeoSeg-PV && pip install -r requirements.txt
python pv_inference.py -i image.tif -w ftunetformer_fom/fold0.safetensors -o output --prob
from huggingface_hub import hf_hub_download
from geoseg.pv import load_model, predict_proba

path = hf_hub_download('nikolukac/GeoSeg-PV-models', 'ftunetformer_fom/fold0.safetensors')
model = load_model(path, device='cuda')
pv = predict_proba(model, rgb, device='cuda') >= 0.5   # rgb: uint8 array (H, W, 3), any size

load_model builds the network from the arch and fom metadata and loads the weights without needing any backbone weights. predict_proba normalises the image, predicts it in 512 × 512 patches and blends overlapping patches. Several models can be passed as a list and are averaged, for example the five splits of one configuration.

Input. 8-bit RGB, normalised with the ImageNet mean and standard deviation (albumentations.Normalize()). The models were trained on Google satellite imagery at zoom level 20 resampled to a ground sampling distance of 0.096 to 0.099 m. Resample other imagery to about 0.1 m (pv_inference.py --gsd 0.1).

Output. Two logits per pixel: class 0 is PV, class 1 background. PV is where the softmax of class 0 is at least 0.5.

Results

F1 and IoU of the PV class on the test tiles of each split (8 tiles, central 512 × 512 pixels each). Every file was re-evaluated after the export and gives exactly the published test result of its split.

Config Network Encoder FOM F1 splits 0 to 4 F1 IoU
unetformer_r18 UNetFormer ResNet-18 no 0.909 / 0.912 / 0.910 / 0.896 / 0.900 0.905 ± 0.007 0.827 ± 0.012
unetformer_r18_fom UNetFormer ResNet-18 yes 0.952 / 0.950 / 0.941 / 0.924 / 0.929 0.939 ± 0.012 0.886 ± 0.022
unetformer_effb5 UNetFormer EfficientNet-B5 no 0.915 / 0.921 / 0.917 / 0.912 / 0.907 0.914 ± 0.005 0.842 ± 0.009
unetformer_effb5_fom UNetFormer EfficientNet-B5 yes 0.958 / 0.951 / 0.947 / 0.938 / 0.946 0.948 ± 0.008 0.901 ± 0.014
unetformer_seresnext26d UNetFormer SE-ResNeXt26d-32x4d no 0.915 / 0.914 / 0.914 / 0.914 / 0.899 0.911 ± 0.007 0.837 ± 0.012
unetformer_seresnext26d_fom UNetFormer SE-ResNeXt26d-32x4d yes 0.955 / 0.954 / 0.949 / 0.940 / 0.913 0.942 ± 0.018 0.891 ± 0.031
ftunetformer FT-UNetFormer Swin-B no 0.919 / 0.922 / 0.914 / 0.914 / 0.905 0.915 ± 0.006 0.843 ± 0.011
ftunetformer_fom FT-UNetFormer Swin-B yes 0.959 / 0.960 / 0.953 / 0.942 / 0.946 0.952 ± 0.008 0.908 ± 0.014
dcswin DCSwin Swin-S no 0.901 / 0.905 / 0.902 / 0.897 / 0.892 0.899 ± 0.005 0.817 ± 0.008
dcswin_fom DCSwin Swin-S yes 0.937 / 0.944 / 0.939 / 0.927 / 0.923 0.934 ± 0.009 0.876 ± 0.015
banet BANet ResT-Lite no 0.943 / 0.939 / 0.935 / 0.916 / 0.901 0.927 ± 0.018 0.864 ± 0.031
banet_fom BANet ResT-Lite yes 0.953 / 0.945 / 0.944 / 0.924 / 0.918 0.937 ± 0.015 0.882 ± 0.026
a2fpn A2-FPN ResNet-18 no 0.895 / 0.903 / 0.904 / 0.896 / 0.892 0.898 ± 0.005 0.815 ± 0.009
a2fpn_fom A2-FPN ResNet-18 yes 0.938 / 0.945 / 0.939 / 0.926 / 0.920 0.934 ± 0.010 0.876 ± 0.018
manet_r18_fom MANet ResNet-18 yes 0.949 / 0.946 / 0.942 / 0.931 / 0.939 0.941 ± 0.007 0.889 ± 0.012
manet_effb5_fom MANet EfficientNet-B5 yes 0.960 / 0.954 / 0.954 / 0.934 / 0.944 0.949 ± 0.010 0.904 ± 0.019
manet_seresnext26d_fom MANet SE-ResNeXt26d-32x4d yes 0.957 / 0.956 / 0.952 / 0.938 / 0.935 0.948 ± 0.010 0.901 ± 0.018

The weights of MANet without FOM were not kept. Its results in the paper are F1 0.922, 0.942 and 0.947 (ResNet-18, EfficientNet-B5, SE-ResNeXt26d-32x4d).

To score a new image against the ground truth, use weights of a split in which that tile was a test tile (see metrics.json or splits.json of the ground truth). The other tiles were seen in training or validation.

Training data and procedure

  • Data. 80 tiles over Slovenia, each with one PV system in its centre, from Google satellite imagery (zoom level 20, downloaded in July 2023). Each tile image was resized to 1024 × 1024 pixels (bicubic) and its central 512 × 512 pixels were used. Modules were annotated with an unsupervised max-tree method on the luminance and corrected by hand. The masks are published separately on Zenodo (doi:10.5281/zenodo.24404853).
  • Splits. Five random splits into 60 training, 12 validation and 8 test tiles (the test sets of different splits may share a few tiles). Training tiles were added flipped horizontally and vertically, and rotated by a random multiple of 90° (probability 0.5) during training.
  • Training. Encoders initialised with ImageNet weights. AdamW (learning rate 1e-4, weight decay 2.5e-4) with Lookahead, cosine annealing with warm restarts (T0 = 30 epochs, minimum 1e-7), batch size 4, at most 200 epochs, early stopping after 20 epochs without improvement of the validation F1. Loss: soft cross-entropy (smoothing 0.05) plus Dice; UNetFormer also with its auxiliary head (weight 0.4). The weights are those of the epoch with the best validation F1.

Citation

@article{lukac2025pv,
  author  = {Luka{\v{c}}, Niko and Seme, Sebastijan and Sreden{\v{s}}ek, Klemen and {\v{S}}tumberger, Gorazd and Mongus, Domen and {\v{Z}}alik, Borut and Bizjak, Marko},
  title   = {Detection and Optimization of Photovoltaic Arrays' Tilt Angles Using Remote Sensing Data},
  journal = {Applied Sciences},
  volume  = {15},
  number  = {7},
  pages   = {3598},
  year    = {2025},
  doi     = {10.3390/app15073598}
}

@article{lukac2013rating,
  author  = {Luka{\v{c}}, Niko and {\v{Z}}laus, Danijel and Seme, Sebastijan and {\v{Z}}alik, Borut and {\v{S}}tumberger, Gorazd},
  title   = {Rating of roofs' surfaces regarding their solar potential and suitability for {PV} systems, based on {LiDAR} data},
  journal = {Applied Energy},
  volume  = {102},
  pages   = {803--812},
  year    = {2013},
  doi     = {10.1016/j.apenergy.2012.08.042}
}

Acknowledgements

University of Maribor, Faculty of Electrical Engineering and Computer Science. Funded by the Slovenian Research and Innovation Agency (ARIS): research project J7-50095 and research core funding P2-0041 and P2-0115. The networks come from GeoSeg by Libo Wang. Satellite imagery: Google, Airbus, CNES/Airbus, Maxar Technologies.

License

The weights and this model card are licensed under CC BY 4.0. Please cite the paper above when you use them. The satellite imagery they were trained on is not part of this release.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support