Applied Statistics Project: Deep Learning for Individual Solar Installation Detection from Aerial Images

Second-year project at ENSAE in partnership with Réseau Transport d'Électricité (RTE).

Grade: 18/20

---

Key Idea

Estimating diffuse electrical production is a growing challenge for forecasting electricity production and consumption in French households. Many households install solar panels without being systematically recorded in administrative data[1]. This lack of exhaustive registration makes it difficult to accurately predict national electricity consumption and optimize electricity distribution (RTE forecasts).

To tackle this issue, we deploy two convolutional neural network models:

• LeNet-5 (TensorFlow)
• ResNet-18 (PyTorch)

These models classify aerial images extracted from Google Earth Engine, labeled by Kasmi et al. (2023)[1].

---

Repository Structure

Jupyter Notebooks

• Stats_desc.ipynb - Descriptive statistics of the Kasmi et al. dataset.
• LeNET5.ipynb - Implementation of the LeNet-5 CNN for detecting solar panels.
• ResNet18-pre_trained.ipynb - Utilizes a pre-trained ResNet-18, retrained on the BDPV database.
• ResNet18-transfer_learning.ipynb - Freezes ResNet-18's convolutional layers and retrains only the fully connected layers on the BDPV database.
• ROC_PR_curve_ResNet18.ipynb - Compares the performance of both ResNet-18 models using ROC and precision-recall curves.
• delete_NA_img.ipynb - Contains a function to remove Google images that include a "gray mask" (equivalent to missing data).

Folders

• src/ - Python modules for the project, including:
  • LeNet-5 architecture
  • Data loader
  • Confusion matrix generation
• weights/ - Pre-trained model weights.
• rapport.pdf - Project report.

---

Execution Guidelines

Running ResNet-18 Pre-trained Model

The ResNet18-pre_trained.ipynb notebook should be executed on the INSEE cloud using the Jupyter-pytorch-gpu service with:

• Minimum Persistence: 15GB
• Dataset Download & Extraction: Kasmi et al. (2023)
• GPU Configuration: Adjust based on the machine setup.

For an example of a properly configured working session, see this link.

---

References

1. A crowdsourced dataset of aerial images with annotated solar photovoltaic arrays and installation metadata - Kasmi et al., 2023
2. Deep Residual Learning for Image Recognition - He et al., 2015