Introduction
WHU-OPT-SAR comprised RGB, near infrared (NIR) optical images and corresponding SAR images, covering an area of 51448.56 km2 with a resolution of 5-meters. As far as we know, WHU-OPT-SAR is the first and largest land use classification dataset that has fused high resolution optical and SAR images with sufficient annotation.
Details
Compared with land cover classification, land use classification focuses more on the social attributes of land use, which can reveal more advanced information of human activities and is closely related to the process of urbanization and socio-economic development. We use the optical images of GaoFen-1 (GF-1) satellite and the SAR images of GaoFen-3 (GF-3) satellite in the same area to build a large-scale optical SAR joint land use classification dataset, named WHU-OPT-SAR.
The WHU-OPT-SAR dataset contains 100 optical images of 5556*3704 pixels and SAR images in the same area, covering an area of about 50000 km2 in Hubei Province (30°N-33°N, 108°E-117°E), China. This area has a subtropical monsoon climate, with lowest altitude of 50m and highest altitude of 3000m. WHU-OPT-SAR covers a wide range of remote sensing images with different terrains such as mountains, woodlands, hills, plains and different vegetation such as coniferous forests, broad-leaved forests, shrubs and aquatic vegetation. Images in this dataset with pixel-level annotations can provide data sources for land use classification based on deep learning. And its well-trained model can be used to training other similar tasks in the remote sensing field.
0->background，10->farmland，20->city，30->village，40->water，50->forest，60->road，70->others。
Link
Links to datasets：https://pan.baidu.com/s/1sIGsD3lBEogSCqzbDOaclA password：i51o
Or
https://drive.google.com/drive/folders/1W3iMpkehng7ADXmhz9pPvmgFQBayq22t?usp=sharing
We have uploaded the prediction results of the dataset to Google Cloud Disk and Baidu Cloud Disk!!!
Statement
In the process of revising the paper, it was found that the proportions of some categories in Table 1 could not correspond to the categories themselves. However, due to the failure of the mailbox, it was not possible to communicate with the editor in time at that time. So the proportion of categories in the published papers is wrong. It is specially stated here that the proportion of each category is based on the actual public dataset.