/solarpvnet

Deep Learning for solar PV ID in aerial images for open New York City data

Primary LanguagePython

SolarPVNet

SolarPVNet is a project to create a neural network that can identify solar PV panels from aerial imagery. The repository contains various tools and codes that have been used to generate data in research articles published by the author.

Data Sources

Data from the following six datasets has been incorporated into existing and
in-progress publications. All datasets require some preprocessing to configure for the data pipeline used by the additional tools. These can be found in the preprocess package. Preprocess scripts do not use a nicely unified format due to the differences between the datasets and will need some individual inspection for how to run them.

CA-F and CA-S - California, USA

These datasets represent the Fresno and Stockton portions of a dataset using USGS orthoimagery from California produced by Bradbury, et al. The dataset additionally contains labels for Modesto and Oxnard, but these were too small for use in the current project (which required 1000+ tiles per site). Images are large, and preprocessing is required. See preprocess_cal.py.

FR-G and FR-I - France

Two datasets by Kasmi, et al. FR-G utilizes data from Google earth, while FR-I uses data from the French national institute of geographical and forestry information (IGN). See preprocess_france.py.

DE-G - Oldenburg, Germany

This dataset was taken from Google Earth imagery of Oldenburg, Germany. The data is not publicly available at present. An initial publication utilizing the data can be found in the proceedings of the IEEE PVSC.

NY-Q - Queens, New York, USA

This dataset consists of all aerial orthoimagery from Queens, New York, for the year 2018. Labeling of the dataset is ongoing, but a public version of the labels is expected to be released in the future. Images are large, and preprocessing is required. See preprocess_nyc.py. Available from: New York Digital Orthodata

Links to other datasets not used at present

Setup Instructions

These tools are written in python using tensorflow,

See dependencies.txt

Install packages

Open a command prompt and type the following

pip install pillow
pip install numpy

Processing Data

Labelling workflow

The workflow for labelling the NYC data is described here.

  • Download data
  • Convert JPEG2000 to PNG
    • convert_to_png.py contains a function to convert a whole directory of files into png format, stripping the alpha channel
  • Label the Data
    • Use labelme. Optionally, use remove_json_imagedata.py to clear the imagedata accidentally saved to json files.
    • Use generate_blank_json.py to fill in any JSON files for images that have no label objects
  • Convert JSON Polygons to Mask
    • Use json_to_dataset.py to convert a directory of the JSON files into binary mask images
  • Slice the Images to Appropriate Size
    • Use split_image.py to slice the images into proper sizes that will be used by the CNN

Installing Labelme

Info found here: LabelMe. I used the prebuilt binaries but you could follow different installation instructions.

Within LabelMe, open the folder where all the PNG files have been saved. Create polygon labels (you will need to zoom in). The polygons should closely hug the PV arrays and should only cover the actual PV panels themselves. Here's an example: Sample Image

Labels used are:

  • pv for obvious pv installations.
  • notpv for empty rings within a larger polygon.
  • maybe for uncertain spots that need review.

Some detailed LabelMe examples here.