Click the blue word vision algorithm above to get more dry goods
In the upper right ... set as a star ★, see you soon
For sharing only, does not represent the position of this official account, the infringement contact will be deleted
Reprinted from: IMARS Remote Sensing Big Data Intelligent Mining and Analysis
The picture comes from the Internet
The segmentation effect of semantic segmentation in natural datasets has been continuously improved, and some studies have gradually applied it to the field of remote sensing, especially high-resolution remote sensing images. Since remote sensing images have the characteristics of massive data, scale dependence, and strong spatial correlation, semantic segmentation methods can be used to extract ground objects or classify them.
With the introduction of fully convolutional neural networks, convolutional networks have not only improved in the classification of full schemas, but also made progress in the local task of structured output. The fully convolutional neural network achieves pixel-level classification of images, thus solving the problem of semantic-level image segmentation.
This issue summarizes the existing 10 remote sensing semantic segmentation datasets, and attaches download links for everyone to test their algorithms.
The shared semantic segmentation datasets are as follows:
1. Gaofen Image Dataset(GID)
2. ISPRS Test Project on Urban Classification and 3D Building Reconstruction—2D SemanticLabeling Contest
3. 2017 IEEE GRSS Data Fusion Contest
4. Aerial Image Segmentation Dataset
5. 2018 IEEE GRSS Data Fusion Contest
6. EvLab-SS Dataset
7. DeepGlobe Land Cover Classification Challenge
8. 38-Cloud dataset
9. Aeroscapes
10. SEN12MS
01
Gaofen Image Dataset(GID)
Gaofen Image Dataset (GID) is a large dataset for land use and land cover (LULC) classification. It contains 150 high-quality Gaofen-2 (GF-2) images from over 60 different cities in China covering a geographic area of more than 50,000 km². GID images have high intra-class diversity and low inter-class separability. GF-2 is the second satellite of the High Definition Earth Observing System (HDEOS). The GF-2 satellite includes panchromatic images with a spatial resolution of 1 m and multispectral images with a spatial resolution of 4 m, with an image size of 6908 × 7300 pixels. Multispectral provides images in blue, green, red, and near-infrared bands. Since its launch in 2014, GF-2 has been used in important applications such as land surveys, environmental monitoring, crop estimation, construction planning, etc.
Download address :
http://captain.whu.edu.cn/GID/
references:
Tong X Y, Xia G S, Lu Q, et al. Learning Transferable Deep Models for Land-Use Classification with High-Resolution Remote Sensing Images[J]. arXiv preprint arXiv:1807.05713, 2018.
02
ISPRS Test Project on Urban Classification and 3D Building Reconstruction—2D Semantic Labeling Contest
ISPRS provides two state-of-the-art airborne image datasets for urban classification and 3D building reconstruction test projects. The dataset employs a digital surface model (DSM) produced from high-resolution orthographic photographs and corresponding dense image matching techniques. Both dataset regions cover urban scenes. Vaihingen is a relatively small village with many individual buildings and small multi-storey buildings; Postdam is a typical historical city with large buildings, narrow streets and dense settlement structures. Each dataset has been manually classified into the 6 most common land cover categories.
① Impervious surface (RGB: 255, 255, 255)
②Building (RGB: 0, 0, 255)
③Low vegetation (RGB: 0, 255, 255)
④Trees (RGB: 0, 255, 0)
⑤Car (RGB: 255, 255, 0)
⑥Background (RGB: 255, 0, 0)
Background classes include bodies of water and objects that are distinct from other defined classes (e.g., containers, tennis courts, swimming pools), which are often uninteresting semantic objects in urban scenes.
Download address :
http://www2.isprs.org/commissions/comm3/wg4/semantic-labeling.html
Vaihingen
The dataset contains 33 remote sensing images of various sizes, each extracted from a larger top-level orthophoto image. The spatial resolution of the top-level image and DSM is 9 cm. The remote sensing image format is an 8-bit TIFF file, which consists of three bands: near-infrared, red and green. DSMs are single-band TIFF files with grayscale levels (corresponding to DSM heights) encoded as 32-bit floating point values.
Postdam
Similar to the Vaihingen region, this dataset also consists of 3-band remote sensing TIFF files and single-band DSM. The coverage size of each remote sensing image is the same. In this way, remote sensing imagery and DSM are defined on the same reference system (UTM WGS84). Each image has an affine transform file to re-decompose the image into smaller pictures when needed.
In addition to DSM, the dataset also provides normalized DSM, i.e., after ground filtering, the ground height of each pixel is removed, resulting in a representation of the height above the terrain. These data were generated using some fully automated filtering workflows, with no manual quality control. Therefore, there is no guarantee that there are no erroneous data here, this is to help researchers use altitude data and not absolute DSM.
03
2017 IEEE GRSS Data Fusion Contest
The 2017 IEEE GRSS Data Fusion Competition adopts classification as the main body. The task to be performed is to classify land use in various urban environments (ie for the 2012 Local Climate Zones (LCZs)). The competition selected several cities to test the ability of the LCZ predictions to be rolled out around the world. The input data is multi-temporal, multi-source and multi-modal, including image and semantic layers.
The dataset consists of 4 parts.
Ground satellite data provided by the US Geological Survey: 8 multispectral bands including visible light, shortwave and longwave infrared, and resampled at 100 m resolution;
Sentinel 2 image: The image has a spatial resolution of 100 m and has 9 multispectral bands, namely visible, near-infrared and short-infrared wavelengths (contains modified Copernicus data 2016);
Secondary data: Open Street Map (OSM) layers with land use information: buildings, nature, roads and other land use areas. It also provides raster maps of OSM layers at 20 m resolution for building and land use areas, which can be overlaid with satellite imagery.
In addition, for selected cities, ground truth labels for various LCZ classes in several regions of the city are also provided. Labels are raster images with 100 m resolution that can also be overlaid onto satellite imagery.
Download address :
http://www.grss-ieee.org/2017-ieee-grss-data-fusion-contest/http://dase.ticinumaerospace.com/index.php
04
Aerial Image Segmentation Dataset
The aerial imagery is split into aerial remote sensing imagery from Google Maps and pixel-level building, road, and background labels from OpenStreetMap. Coverage areas are Berlin, Chicago, Paris, Potsdam and Zurich. The ground-truth imagery includes an aerial image of the Tokyo area from Google Maps, as well as manually generated, pixel-level labels for buildings, roads, and backgrounds. Pixel-level labels are provided as PNG images in RGB order, and pixels labeled buildings, roads, and backgrounds are represented by RGB colors [255, 0, 0], [0, 0, 255], and [255, 255, 255].
Download address :
https://zenodo.org/record/1154821#.XH6HtygzbIU
05
2018 IEEE GRSS Data Fusion Contest
Data were obtained by NCALM on 16 February 2017 between 16:31 and 18:18 GMT from the National Center for Airborne Laser Mapping. The sensors used for data collection in this competition include: LiDAR sensor OPTech TITAM M (14sen/con340) with 3 different bands, Dimac ULTRALIGHT+ high resolution color imager with 70 mm focal length, Hyperspectral imager ITRES CASI 1500. The multispectral lidar point cloud data bands are at 1550 nm, 1064 nm and 532 nm. The hyperspectral data covers the range of 380-1050 nm, with a total of 48 bands and a spatial resolution of 1 m. The high-resolution RGB remote sensing images have a spatial resolution of 5 cm and are segmented into several individual pictures.
Download address :
http://www.grss-ieee.org/community/technical-committees/data-fusion/2018-ieee-grss-data-fusion-contest/
http://dase.ticinumaerospace.com
06
EvLab-SS Dataset
The EvLab-SS dataset is used to evaluate semantic segmentation algorithms on real engineering scenes, aiming to find a good deep learning architecture for high-resolution pixel-level classification tasks in the remote sensing domain.
The dataset is derived from the China Geographic Condition Survey and Mapping Project, and each image is fully annotated with the Geographic Condition Survey. The average size of the dataset is about 4500 × 4500 pixels. The EvLab-SS dataset contains 11 broad categories, namely background, farmland, garden, woodland, grass, building, road, structure, bored pile, desert, and water, and currently includes 60 images captured by different platforms and sensors.
The dataset includes 35 satellite images, of which 19 were acquired by the World-View-2 satellite, 5 by the GeoEye satellite, 5 by the Quick Bird satellite, and 6 by the GF-2 satellite. The dataset also has 25 aerial images, 10 of which have a spatial resolution of 0.25 m and 15 of which have a spatial resolution of 0.1 m.
Download address :
http://earthvisionlab.whu.edu.cn/zm/SemanticSegmentation/index.html
references:
Zhang M, Hu X, Zhao L, et al. Learning dual multi-scale manifold ranking for semantic segmentation of high-resolution images[J]. Remote Sensing, 2017, 9(5): 500
07
DeepGlobe Land Cover Classification Challenge
The DeepGlobe Land Cover Classification Challenge is a public dataset that provides high-resolution sub-meter satellite imagery with a focus on rural areas. This dataset is challenging due to the diversity of land cover types and the high density of annotations. The dataset contains a total of 10,146 satellite images with a size of 20,448 × 20,448 pixels and is divided into training/validation/test sets, each with 803/171/172 images (corresponding to 70%/15%/15%).
Download address :
http://deepglobe.org/index.html
https://competitions.codalab.org/competitions/18468
08
38-Cloud dataset
The dataset contains 38 Landsat 8 scene images and their manually extracted pixel-level ground-truth labels for cloud detection. The entire images of these scenes are cropped into 384 × 384 patches to be suitable for deep learning based semantic segmentation algorithms. The dataset is divided into 8400 patches for training and 9201 patches for testing. Each image has 4 corresponding bands, red (band 4), green (band 3), blue (band 2), and near-infrared (band 5).
Download address :
https://github.com/SorourMo/38-Cloud-A-Cloud-Segmentation-Dataset
References :
Mohajerani S, Krammer T A, Saeedi P. Cloud Detection Algorithm for Remote Sensing Images Using Fully Convolutional Neural Networks[J]. arXiv preprint arXiv:1810.05782, 2018
Mohajerani S, Saeedi P. Cloud-Net: An end-to-end Cloud Detection Algorithm for Landsat 8 Imagery[J]. arXiv preprint arXiv:1901.10077, 2019
09
Aeroscapes
The aerial semantic segmentation dataset includes images captured using commercial drones from heights ranging from 5 to 50 m. The dataset provides 3269 720p images and ground truth labels for 11 classes.
download link:
https://github.com/ishann/aeroscapes
references:
Nigam I, Huang C, Ramanan D. Ensemble knowledge transfer for semantic segmentation[C]//2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, 2018: 1499-1508
10
SEN12MS
SEN12MS is a dataset composed of 180,748 corresponding 3 types of remote sensing data, including Sentinel-1 bipolar SAR data, Sentinel-2 multispectral images and MODIS land cover maps. The Sentinel-1 image resolution is 20 m, the Sentinel-2 multispectral image resolution is 10 m, the number of bands is 13, and the land cover image resolution of MODIS is 500 m.
download link:
https://mediatum.ub.tum.de/1474000
end
Click to see support