DeepGeoStat WP5 Solar Panel Trained Network

This dataset contains the weights of a convolutional neural network (CNN) trained to recognize the presence of solar panels on aerial photos. In particular, it contains the saved state of a ResNet50 CNN that has been trained on a dataset containing annotated high-resolution aerial images of two regions in the south of the Netherlands. Many photos in this dataset have been annotated multiple times, and the annotations are not always unanimous. The dataset of aerial images together with annotations can be downloaded from here.

The model for detecting whether solar panels are present in aerial photos has been developed under the DeepSolaris and DeepGeoStat projects. Corresponding Pytorch code can be found here. The code also demonstrates how to load the saved state into a ResNet50 model, and use it for detecting solar panels on aerial photos.

This research was conducted under:

ESS action 'Merging Geostatistics and Geospatial Information in Member States' (grant agreement no.: 08143.2017.001-2017.408),

 ESS topic B5674-2020-GEOS (project 101033951 2020-NL-GEOS-DEEP-GEO-STAT),

a research program of Statistics Netherlands (https://www.cbs.nl)