Background

This project is a pytorch project of the algorithm FCD-GAN. Fully Convolutional Change Detection Framework with Generative Adversarial Network (FCD-GAN) is a newly proposed framework for change detection in multi-temporal remote sensing images. The corresponding publication can be found in the following citation:

C. Wu, B. Du, and L. Zhang, “Fully Convolutional Change Detection Framework with Generative Adversarial Network for Unsupervised, Weakly Supervised and Regional Supervised Change Detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1-15, 2023.

In this project, three demo codes are released for unsupervised, weakly supervised and regional supervised change detection, as "Demo_USSS", "Demo_WSSS", and "Demo_RSSS".

Install

This project uses 'gdal', 'numpy', 'tqdm', 'cv2', 'pil', 'skimage'. Go check them out if you don't have them locally installed.

Usage

The implementations of the algorithms are introduced in this section.

Demo_USSS

This code is a demo for the algorithm of unsupervised change detection. The user should input the bi-temporal images with the following code:

dir = r'/data'
ImageXName = 'T1.tif'
ImageYName = 'T2.tif'
RefName = 'ref.tif'·

In order to customize the reference, the user can define the value to indicate change or nonchange in the reference image with the following code:

gt_map = [1, 2]

This code will output two images (the color image is optional) and one txt file. The first image is a change density image, with the range of [0, 1]. Higher value indicates higher probability to be changed. This image will be outputed in the following path:

outdir = dir
ext = '_l1w065_pw04_github'
CMapName = 'ChangeDensity{}'.format(ext)

The second image is a colorful evaluation image, where the value of {0, 1, 2, 3} indicates {FP, FN, TP, TN}. The density image is firstly assigned to be unchanged (0) or changed (1) with the given prob_threshold in the code, and evaluated with the reference to produce the colorful evaluation image is optional with the following switch:

write_color = True

The colorful evaluation image will be outputed in the following path:

OutColorPath = os.path.join(outdir, "{}_acc_color{}".format(CMapName, ext1))

The last output file is a txt file, which will record the parameter settings of this code. The txt file will be outputed in the following path:

ParaTxtPath = os.path.join(outdir,'Para_{}{}.txt'.format(time.strftime("%b%d%H%M", time.localtime()), ext))

Demo_RSSS

This code is a demo for the algorithm of regional supervised change detection, which means the changed areas are labelled with probable regions. In the demo code, the experimental data is the pre-processed OSCD data. The users can customize their experimental data with pytorch. The core of the dataset is to read the multi-temporal images and the supervised regions (0: unchanged; 1: changed). Most inputs and outputs are similar with Demo_USSS. The notice is: The generator model trained can be saved and reused in the following path:

OutGModelDir = r'/GModel'

Or you can also train a generator model in each run if you turn off the following switch:

modelG_reuse = True

The density images and colorful images are outputed in each image file with the following extName:

outName_density = 'density'
outName_binary = 'color'

The discrimitor model and segmentor model can be caved in the following path:

extName = '_l1002_r2_d1_g05_github'
OutDir = os.path.join(imgDir, 'model{}'.format(extName))

About the Dataset

Demo_USSS can diretly process the whole multi-temporal remote sensing images with the format of '.tif'. Demo_WSSS need to read the image with a specific format. The WHU Building dataset can be transformed into a weakly supervised change detection with the code 'BuildingProcess.py'