Repository files navigation

The number of photovoltaic power plants is growing so rapidly that we must rely on satellitel observations and efficient machine learning methods for the global monitoring.MUSIC for P3 (PhotovoltaicPowerPlants) is a training and validation dataset generated byAIRC/AIST to support such a global survey of photovoltaic power plants.

We have picked up the photovoltaic power plants listed onthe website of Electrical Japan only if they generated more than 5MW electricity and the construction had been completed by 2015. The multiband satellite images of these target areas, taken byLandsat-8, were cropped into a 16 × 16 grid covering a 480 × 480 meter area as shown below.

These patch images are classified as “positives” if the solar panels cover more than 20% of the total areas, while patches with no solar panels are classified as “negatives”. The rest with the intermediate coverage (0~20%) were neither “positives” and “negatives”.

You can download theMUSIC for P3 dataset with two different format (HDF5 andGeoTiff) along with the source code for the detection and classification. More detailed exaplanations can be found in the following papers.

[1]Tomohiro Ishii, Edgar Simo-Serra, Satoshi Iizuka, Yoshihiko Mochizuki, Akihiro Sugimoto, Ryosuke Nakamura, Hiroshi Ishikawa ,"Detection by Classification of Buildings in Multispectral Satellite Imagery," ICPR 2016. (http://www.f.waseda.jp/hfs/IshiiICPR2016.pdf)

[2]Nevrez Imamoglu, Motoki Kimura, Hiroki Miyamoto, Aito Fujita, Ryosuke Nakamura,"Solar Power Plant Detection on Multi-Spectral Satellite Imagery using Weakly-Supervised CNN with Feedback Features and m-PCNN Fusion," BMVC 2017. (https://arxiv.org/abs/1704.06410)

It should be noted here that the "negatives" in the initial dataset (V1) is contaminated by small-scale photovoltaic power plants not included in theoriginal inventory. After publishing these papers, we have checked all the false positives through in-site survey or high-resolution imagery. In fact, some of "false positives" are found to be relatievely small photovoltaic power plants unlisted in the original database. Therefore, we've produced updated V2 dataset by correcting these misclassification. The performance of our previous works was estimated with V1 dataset, but more accurate estimate can be achieved by using this updated dataset V2.

IMPORTANT -- Please read theTerms of Use before downloading the MUSIC4P3 dataset.

HDF5 version (for usingTorch version code) of the dataset can be downloaded fromhere (4.6GB) .
Or type the following in the terminal.

$ wget http://data.airc.aist.go.jp/MUSIC4P3dataset/MUSIC4P3data_hdf.zip$ unzip MUSIC4P3data_hdf.zip

The directory configuration in the unzipped folder:

./torch/resource/train/  LC81060302015147LGN00.hdf5LC81070302015266LGN00.hdf5 ...val/LC81060302015147LGN00.hdf5LC81070302015266LGN00.hdf5 ...test/LC81060302015147LGN00.hdf5LC81070302015266LGN00.hdf5 ...

V2 dataset ishere.

Tiff version (for usingChainer version code) of the dataset can be downloaded fromhere (4.2GB) .
Or type the following in the terminal.

$ wget http://data.airc.aist.go.jp/MUSIC4P3dataset/MUSIC4P3data_tiff.zip$ unzip MUSIC4P3data_tiff.zip

The directory configuration in the unzipped files is as follows:

./chainer/resource/train/positive/LC81060302015147LGN00_274_300_2064.tiffLC81060302015147LGN00_274_300_13289.tiff ...negative/LC81060302015147LGN00_100_105.tiffLC81060302015147LGN00_100_114.tiff ...val/positive/LC81070352015298LGN00_206_266.tiffLC81070352015298LGN00_374_160.tiff ...negative/LC81060302015147LGN00_2_86.tiffLC81060302015147LGN00_4_88.tiff ...test/positive/LC81060302015147LGN00_79_323.tiffLC81060302015147LGN00_80_323.tiff ...negative/LC81060302015147LGN00_100_100.tiffLC81060302015147LGN00_100_101.tiff ...

V2 dataset ishere.

• torch7
  For torch installition seehttp://torch.ch/
• torch_toolbox
  Download fromhttps://github.com/e-lab/torch-toolbox
  Unzip and put the it on the same directory.

FOR EXAMPLE<dir>torch/<dir>resource<dir>torch-toolboxtrain-cnn.luatrain-cnn.shtest-cnn.luatest-cnn.sh ...

For trainging, type the followingin the terminal.

$ sh train-cnn.sh

Models are saved to "./ishiinet_p1n15/cnn_nm_epXXXX.net" (the output file name is specified in train-cnn.sh) .

Type the following in the terminal.

$ sh test-cnn.sh

Result file are saved to "./ishiinet_p1n15/test_th0.999_all_cm.txt".
Can be changed by passing in this shell file.
If you want to the same result from paper[1], please use thispre training model.

• Python 2.7.*
• Chainer 1.24.*
  For chainer Installition, see fromhttps://chainer.org/

Type the following in the terminal.

$ sh train.sh

Models are save to ./result/ (the output directory is specified in train.sh) .

Type the following in the terminal.

$ sh test.sh > result.log

Result file are created to ./result.log .
If you want to the same result from paper[1], please use thispre training model.

This dataset and source code are based on results obtained from a project commissioned by the New Energy and Industrial Technology Development Organization (NEDO).