The méthode is based on the IGARSS paper Oil tank detection in satellite images via a contrario clustering.

From the B02, B03, B04 and B08 bands, a greyscale image is computed via PCA. Then the Canny-Devernay procedure is applied to detect closed edge segments on which an isoperimetric threshold is applied.

This gives a first detection of tanks which are then clustered by using binary dilation of varius radii. Finally, the clusters are filtered based on their NFA score.

To use the code, you must first compile the Canny-Devernay edge detector code:

mkdir build
cd build
cmake ../devernay1.1/
make

The binary files will be moved to TankDetection/C_exe.

The detector function is named detect_tank and is in TankDetection/detect_tanks.py. The file TankDetection/build_crops.py contains a wrapper for sentinel-2 .jp2 images.

The options for this wrapper are the following :

  -h, --help                	show this help message and exit
      --input INPUT         	folder that all the input images are stored
      --output OUTPUT       	output path
      --iso_th ISO_TH       	isoperimetric threshold
      --zoom                	apply 2x fft_zoom
      --save_res            	save result images  
      --auto_th             	use Otsu's method to set Canny-Devernay's threshold
  -lth LOW, --low LOW  			low threshold for Canny-Devernay
  -hth HIGH, --high HIGH 		high threshold for Canny-Devernay
  -cw CROP_W, --crop_w CROP_W 	Crop width
  -ch CROP_H, --crop_h CROP_H  	Crop height

Usage :

python build_crops.py --input ./im_test --output ./test_result --save_res --auto_th --zoom

build_crops will partition the input image (.jp2 or .tif files) according to the -cw and -ch options. Each crop will then be processed to detect tank farm and the results will be stored in the folder specified by the --output argument, were a folder will be named after the chosen parameters. In this file you will find a folder for each crop with the gray_scale image, the edge map, and the json file where the detections are saved. --save_res will save the image result for analysis.