Cloud-Removal-in-Satellite-Images-using-Conditional-Generative-Adversarial-Networks

Affiliation

Photogrammetry and Remote Sensing Department, Indian Institute of Remote Sensing, ISRO, Dehradun

April 2020 - July 2020

Summary

Cloud cover in the earth's atmosphere is a major issue in temporal optical satellite image processing. Clouds, thick or thin, cover the earth features in a satellite image and hide important information. The objective of this project is to develop a deep learning based automated pipeline to remove cloud cover from optical satellite imagery. Conditional GANs architecture is used to learn the mapping of the cloudy satellite image to it's cloud-free counterpart. A novel augmented and computationally efficient training approach is suggested.

Modules

Built With/Libraries Used/Network Architectures

Keras (TF Backend)
Numpy
Matplotlib
PatchGAN
U-Net

Dataset

Sentinel-2 is a part of the opensource earth-observation Copernicus mission of the ESA. It provides high temporal resolution multispectral data (13 bands) at 10m spatial resolution. False Color Composite Images (Green, Red and NIR bands) are used. The study site is chosen based availability of phenological features (crops/agriculture). Removing cloud cover in Sentinel-2 satellite images using only optical data and a novel augmented training approach using conditional GANs.

Currently the training dataset is augmented from a single pair of cloudy and cloud-free images labelled as True and False image and used as Training and Test Data respectively.

The initial augmentation is performed only by rotation of 0, 90, 180 and -90 degrees which gives us 4 pairs, say 1,2,3 and 4.These four pairs are then stacked iteratively 10 times as: [1,2,3,4,1,2,3,4,1,2]. Later on, more data augmentation was implemented by performing skew operations on the cloudy and cloud-free images, and then perform the rotation operations. Now there are 4 pairs of original (unskewed) images, say, [1, 2, 3, 4] and 4 pairs of the skewed images, say [1', 2', 3', 4'] which are stacked as [1,2,3,4,1',2',3',4',1,1'].

Training Data Drive Link

Skew Augmented Training Data Drive Link

Model and Training

pix-2-pix GAN model is used for learning the mapping of the cloudy image to it's cloud-free counterpart. It follows a supervised conditional vector based training approach. The model is trained on the cloudy and cloud-free image pairs which are dated two days apart. pix-2-pix model follows a pixel to pixel image restoration approach and generates a new pixel cloud-free for every cloudy pixel.

Trained Model Drive Link

Architectures

Discriminator Architecture (PatchGAN)

Generator Architecture (U-Net)

Model Performance and Analyzing results on Skewed Dataset

Various Parameters for evaluation per 100 epochs are as follows in terms of PC,PSNR and SSIM on train and validation data:

Epoch	NIR_PC_train	NIR_PC_val	Red_PC_train	Red_PC_val	Green_PC_train	Green_PC_val	PSNR_train	PSNR_val	SSIM_train	SSIM_val
100	-0.05000842315	-0.0883398788	-0.014429011	-0.01942083916	0.155553528	0.1314745827	10.5787193	9.944368826	0.1558209028	0.1387626019
200	-0.09461359599	-0.1647732402	0.09893491829	0.1078286654	0.07439906456	0.06439988178	19.49975022	17.96775821	0.5434919468	0.4879971859
300	0.4844487891	0.04423792917	-0.202653834	0.2110942252	0.100536630	0.1016198901	27.39418744	24.93884501	0.9066141119	0.8419657794
400	0.5659931852	0.1553301583	0.3359637375	0.2660727986	0.2462063361	0.1102983506	31.2101439	27.2062056	0.9494872523	0.8889470705
500	0.1064666356	0.0688902821	0.00386225793	0.000663448184	-0.06814932682	-0.1012978036	20.8824697	19.21863675	0.4620478366	0.4057640266
600	0.242182867	0.1469791003	0.1299660457	0.1374295475	-0.02266885263	-0.05183500483	24.81609684	22.2485777	0.6618107623	0.6026207847
700	0.1141179129	0.09077203119	0.0231577498	0.0461014954	0.01413759484	0.01533657492	15.81737335	14.91596596	0.1561151655	0.1371105281
800	0.07355812207	0.07218999338	0.03420086299	0.05645767604	0.04626819331	0.05379545084	14.91536278	14.151824	0.1232767038	0.1109237418
900	0.1402173311	0.08612228292	0.08809589865	0.0939511939	0.1013317205	0.09384005917	18.03352817	16.77826748	0.2319571403	0.2019726696
1000	0.1522175639	0.03738729948	0.1245654264	0.1658647433	0.1536526258	0.1253755692	22.9342766	20.83937953	0.489397219	0.4237837526
1100	0.2207068608	0.08419658372	0.1730546559	0.2173727323	0.219995931	0.1654665957	26.96671849	24.0358739	0.7026804423	0.6182513077
1200	0.2352971555	0.0835442674	0.1249354459	0.147639293	0.1379616916	0.1153903569	26.10884684	23.56929742	0.6732339429	0.5950390088
1300	0.1791357889	0.08168209727	0.008672194105	0.03997902222	0.0101701286	0.006031576638	23.91331973	22.03335297	0.5194873822	0.4578776459
1400	0.3293007461	0.1246761137	0.02887394793	0.05673247339	0.04255831788	0.01574076197	23.80483363	21.73766691	0.5543500455	0.4823385855
1500	0.3038193774	0.06004512987	-0.001345626278	0.03361086924	0.00718772725	0.004563553405	23.92568466	21.85922651	0.6341453641	0.5565858715
1600	0.5292296154	0.1225696892	0.03089674103	0.07625537081	0.08322336163	0.06492719505	28.1318743	24.72636908	0.7972798953	0.7104326953
1700	0.632629049	0.1635108239	0.08676428243	0.1230198421	0.2582400218	0.1075018241	31.25227794	26.33953686	0.8842468747	0.7974064016

Spikes in Skew Dataset Losses

The spikes in the generator loss graph can be due to a few reasons such as the image dataset might have a few big contrasts due to augmenting the dataset and the Generator might have been stuck in a mode collapse(local minima) which when resolved resulted in very high spikes in the loss as the descriminator was quite good trained till then.

Comparison of Result on the two datasets

Correlation


PC on original data	PC on skewed data

SSIM

PSNR

Contribution

Please feel free to raise issues and fix any existing ones. Further details can be found in our code of conduct.

While making a PR, please make sure you:

Always start your PR description with "Fixes #issue_number", if you're fixing an issue.
Briefly mention the purpose of the PR, along with the tools/libraries you have used. It would be great if you could be version specific.
Briefly mention what logic you used to implement the changes/upgrades.
Provide in-code review comments on GitHub to highlight specific LOC if deemed necessary.
Please provide snapshots if deemed necessary.
Update readme if required.

Credits

This study available as a book chapter. DOI

Please cite this work as:

Maniyar, C., & Kumar, A. (2021). Generative Adversarial Network for Cloud Removal from Optical Temporal Satellite Imagery. In Soft Computing for Problem Solving (pp. 481-491). Springer, Singapore.

Chintan2108/Cloud-Removal-in-Satellite-Images-using-Conditional-Generative-Adversarial-Networks