The goal of this project is to detect solar panels in satellite images using deep learning. Our model is based on U-net and trained on satellite image from the USA and France. In addition to creating a model able to accurately detect solar panels we also wanted to test the model's ability to generalize to other geographic areas. The model performs semantic segmentation, meaning that it predicts the presence of a solar panel for every pixel in the input image.

To configure an environment to run this project we recommend using conda. A requirements.txt is included in the repository. To set up the conda environment, execute the following lines:

conda create -n solar-detection python=3.9
conda activate solar-detection
pip install -r requirements.txt 

The two main modules are data_generation.py handling the generation and processing of the data, and train_model.py for the training of the model.

The data generator class to use depends on the data source, as US and French satellite data have different sizes, and thus require different approaches. The US dataset is composed of images that are 5000 by 5000 pixels in sixe and usage of the DatasetGenerator class is created with images of this size in mind. The French satellite images we had access to were 25 000 by 25000 pixels, and too large to load the whole image in memory. For this reason, a second data generation class, DatasetGeneratorGERS, was created by adapting the data generator class for the US images. In the case of large image samples like the French data the gers argument needs to be set to true, otherwise the script assumes the dataset to be composed of smaller satellite images.

To generate a dataset of image samples and their corresponding binary mask ('0' if there is no panel, '1' if there is) the following arguments can be specified:

• image_path: Path to the satellite images (str)
• json_path: Path to JSON file containing the panel polygons (str)
• dataset_path: Path to where the dataset will be created (str)
• gers: Set to true if generating from Gers data (Bool)
• image_size: Generated image sample size, samples will be image_size by image_size pixels in size (int)
• shuffle: Shuffle after generating samples (Bool)
• test_split: Ratio of samples in the test set (Float)

These arguments can be modified in the datagen_config.json file.

To generate a dataset from satellite images and a polygon file, run the data_generation.py script by executing the following line:

python data_generation.py

Data augmentation is used in the project to get the most out of the accessible dataset. The methods used are vertical and horizontal flips of image samples and its corresponding binary mask. The image sample has a 50% chance of being flipped along each axis, essentially making it possible to generate four total training samples from each training sample in the original dataset. The ground truth label is flipped along with the sample.

A new model can be trained by running the train_unet.py script. When running the script, two arguments can be provided:

• model_name: If in training mode: this will be the name of the saved model. If in evaluation mode, this is the name of the model to be evaluated.
• mode: Can be either t/train or e/eval

The following line shows an example of unsing the train_unet.py script:

python train_unet.py --model_name my_model --mode t

The parameters for the training of the model will be loaded from the train_config.json file. The trained model will be saved in a sub-directory (named after the model) of trained_models.

The dataset generation is adapted for a dataset of satellite imagery with a corresponding JSON file containing the solar panel polygons, and this JSON file has a pre-defined structure that should be followed to generate new samples and ground truth labels without needing to modify the data_generation.py script.
The structure of the JSON file is given below:

{
    "geometry": {
        "coordinates": [
            [-119.83998,36.92599116],
            ...
        ]
    },
    "properties": {
        "polygon_id": 1,
        "image_name": "11ska460890",
        "datum": "NAD83",
        "projection_zone": "11",
        "resolution": 0.3,
        "city": "Fresno",
        "nw_corner_of_image_latitude": 36.92633611,
        "nw_corner_of_image_longitude": -119.8516222,
        "se_corner_of_image_latitude": 36.91323333,
        "se_corner_of_image_longitude": -119.8343,
        "centroid_latitude_pixels": 107.6184581,
        "centroid_longitude_pixels": 3286.151487,
        "polygon_vertices_pixels": [
            [3360.495069,131.6311637],
            ...
        ]
    }
}

Adding more transformation than randomly flipping the image samples along its horizontal and vertical axes is likely to make the model more robust. We would recommend implementing rotations and zooms as a next step in improving data augmentation but would refrain from using transformations like shears and elastic deformations to preserve the characteristic shape of solar panels in the training set.