kd1510/panel_segmentation

Solar Site CV for peak capacity estimation

Contents

1. Introduction
2. Downloading the dataset for fine-tuning
3. Fine-tuning the model
4. Inference

Introduction

• Estimating the total capacity of a solar site from imagery is a common and challenging task.
• The approach taken here is based on semantic segmentation of the site into panels and non-panels to obtain an estimate of total panel surface area.
• Recent advances in transformer-based segmentation models allow us to have few-shot learning capabilities, which is useful for this task due to the scarcity of annotated data. We therefore use a pre-trained SegNet transformer (https://arxiv.org/abs/1511.00561) and fine-tune it on a small dataset of annotated images obtained from a public research initiative.
• Additionally, due to the expensive nature of acquiring satellite imagery, workflows should be able to leverage open datasets such as google maps imagery. Our results show that we can apply few-shot learning to this task and generalize to unseen lower-resolution imagery.

Dataset details

• Images for fine-tuning on sites in various environments can be found at https://zenodo.org/record/5171712

• The following categories of site environments are available:

    - Cropland
    - Grassland
    - Saline/Alkali
    - Shrubwood
    - Water surface
    - Rooftop
  

• Dataset was produced for public use courtesy of Jiang Hou, Yao Ling, & Liu Yujun. (Multi-resolution dataset for photovoltaic panel segmentation from satellite and aerial imagery 2021)

Fine-tuning

• In order to run the training procedure:
  1. Download the portion of the dataset you wish to use for fine-tuning and set the location to the environment variable IMAGE_FOLDER_PATH
  2. Run python train.py to begin training. The model will checkpoints will be saved in the parent directory.

Inference

• To run inference, run MODEL_CHKPT=<path to model checkpoint> TEST_IMAGEPATH=<path to test image> python inference.py