SoilingNet: Convolutional Neural Networks for Analysis of Soiling Defects in Photovoltaic Panels

In our study, we present a two-step, fully supervised deep learning approach for analyzing solar panel soiling on a per-panel level.

Solar panel soiling is an important problem for the energy sector as we transition away from fossil fuels. Previous methods have aimed to analyze and combat soiling through the use of classical computer vision techniques or weakly supervised deep learning. In our study, we present a two-step, fully supervised deep learning approach for analyzing solar panel soiling on a per-panel level. From a single RGB image of a solar panel, combined with the environmental factor of solar irradiance, our method produces predictions for soiling type, soiling location, and soiling severity. We also introduce a first-of-its-kind dataset labeled with ground truth semantic segmentation maps for the purpose of solar panel soiling analysis. This dataset contains 1104 samples. We find that our model can achieve Jaccard indices in excess of 70 when predicting semantic segmentation maps and top-1 accuracy’s in excess of 92% when predicting soiling severity. Our method outperforms previous techniques in the domain of solar panel soiling analysis.

Explore More!

Source Code

Insect Identification Project for Agricultural Advancement

Through the use of image data, we developed an artificial intelligence system which is able to predict an insect’s species based on a photo of a given insect.

Archith S.

Mentored by Barbie Duckworth

workspace_premium

Towards Monarch Butterfly Preservation: Applications of AI in Species Identification

This study developed a deep-learning model to distinguish monarch butterflies from visually similar species, achieving 83.5% accuracy on the testing set. The model could help novice conservationists identify monarchs more efficiently and support monitoring and conservation efforts.