The METER-ML dataset is a multi-sensor Earth observation dataset containing georeferenced images in the U.S. labeled for the presence or absence of six methane source facilities. More information about how the dataset was constructed can be found at:

The project website

The CDCEO 2022 Workshop publication (please cite this paper when citing the dataset)

This dataset consists of 85,066 train, 515 validation, and 1,018 test samples, each containing images from NAIP, Sentinel-1, and Sentinel-2. The folder of train images is split into three parts due to its size; you will need to combine them after downloading. The format of each sample is as follows:

train_dataset/ [latitude]_[longitude]/ naip.png sentinel-1.npy sentinel-2-10m.npy sentinel-2-20m.npy sentinel-2-60m.npy

The NAIP image is stored as a 4-channel PNG image with the NIR band in the alpha channel. The other images are stored directly as NumPy arrays. The channels in each image are in the following order:

sentinel-1: VV, VH sentinel-2-10m: red, green, blue, NIR sentinel-2-20m: RE1, RE2, RE3, RE4, SWIR1, SWIR2 sentinel-2-60m: coastal aerosol, water vapor, cirrus

The labels are found in the corresponding GeoJSON file for each dataset (easily loaded with geopandas), which contains the following columns:

Latitude: latitude coordinate of the image center Longitude: longitude coordinate of the image center Type: label of facility or facilities present in the image Source: data source the coordinates originally came from Image_Folder: folder in the dataset where the image can be found geometry: bounding box for the area covered by the image

If you have questions about the dataset, contact us at:

bwzhu@cs.stanford.edu, niclui@stanford.edu, jirvin16@cs.stanford.edu

"Synthropo" is synthetic dataset created on Nvidia Omniverse Replicator for human body segmentation tasks that takes into account clothing worn by human. Input image is human with various types of clothing. And output is alpha channel of the input images the human without clothing. For convenience, the images has been converted to numpy files.

This database, named UIBVFEDPlus-Light, is an extension of the previously published UIBVFED virtual facial expression dataset. It includes 100 characters, four lighting configurations and 13200 images. Images are in png format with a resolution of 1080x1920 RGB, without alpha channel and an average size of 2.0 MB.The images represent virtual characters reproducing FACS-based facial expressions. Expressions are classified based on the six universal emotions (Anger, Disgust, Fear, Joy, Sadness, and Surprise) labeled according to Faigin’s classification.The dataset aims to give researchers access to data they may use to support their research and generate new knowledge. In particular, to study the effect of lighting conditions in the fields of facial expression and emotion recognition.

This dataset contains both large (A0) printable maps of the Torres Strait broken into six overlapping regions, based on a clear sky, clear water composite Sentinel 2 composite imagery and the imagery used to create these maps. These maps show satellite imagery of the region, overlaid with reef and island boundaries and names. Not all features are named, just the more prominent features. This also includes a vector map of Ashmore Reef and Boot Reef in Coral Sea as these were used in the same discussions that these maps were developed for. The map of Ashmore Reef includes the atoll platform, reef boundaries and depth polygons for 5 m and 10 m.

This dataset contains all working files used in the development of these maps. This includes all a copy of all the source datasets and all derived satellite image tiles and QGIS files used to create the maps. This includes cloud free Sentinel 2 composite imagery of the Torres Strait region with alpha blended edges to allow the creation of a smooth high resolution basemap of the region.

The base imagery is similar to the older base imagery dataset: Torres Strait clear sky, clear water Landsat 5 satellite composite (NERP TE 13.1 eAtlas, AIMS, source: NASA).

Most of the imagery in the composite imagery from 2017 - 2021.

Method: The Sentinel 2 basemap was produced by processing imagery from the World_AIMS_Marine-satellite-imagery dataset (not yet published) for the Torres Strait region. The TrueColour imagery for the scenes covering the mapped area were downloaded. Both the reference 1 imagery (R1) and reference 2 imagery (R2) was copied for processing. R1 imagery contains the lowest noise, most cloud free imagery, while R2 contains the next best set of imagery. Both R1 and R2 are typically composite images from multiple dates.

The R2 images were selectively blended using manually created masks with the R1 images. This was done to get the best combination of both images and typically resulted in a reduction in some of the cloud artefacts in the R1 images. The mask creation and previewing of the blending was performed in Photoshop. The created masks were saved in 01-data/R2-R1-masks. To help with the blending of neighbouring images a feathered alpha channel was added to the imagery. The processing of the merging (using the masks) and the creation of the feathered borders on the images was performed using a Python script (src/local/03-merge-R2-R1-images.py) using the Pillow library and GDAL. The neighbouring image blending mask was created by applying a blurring of the original hard image mask. This allowed neighbouring image tiles to merge together.

The imagery and reference datasets (reef boundaries, EEZ) were loaded into QGIS for the creation of the printable maps.

To optimise the matching of the resulting map slight brightness adjustments were applied to each scene tile to match its neighbours. This was done in the setup of each image in QGIS. This adjustment was imperfect as each tile was made from a different combinations of days (to remove clouds) resulting in each scene having a different tonal gradients across the scene then its neighbours. Additionally Sentinel 2 has slight stripes (at 13 degrees off the vertical) due to the swath of each sensor having a slight sensitivity difference. This effect was uncorrected in this imagery.

Single merged composite GeoTiff: The image tiles with alpha blended edges work well in QGIS, but not in ArcGIS Pro. To allow this imagery to be used across tools that don't support the alpha blending we merged and flattened the tiles into a single large GeoTiff with no alpha channel. This was done by rendering the map created in QGIS into a single large image. This was done in multiple steps to make the process manageable.

The rendered map was cut into twenty 1 x 1 degree georeferenced PNG images using the Atlas feature of QGIS. This process baked in the alpha blending across neighbouring Sentinel 2 scenes. The PNG images were then merged back into a large GeoTiff image using GDAL (via QGIS), removing the alpha channel. The brightness of the image was adjusted so that the darkest pixels in the image were 1, saving the value 0 for nodata masking and the boundary was clipped, using a polygon boundary, to trim off the outer feathering. The image was then optimised for performance by using internal tiling and adding overviews. A full breakdown of these steps is provided in the README.md in the 'Browse and download all data files' link.

The merged final image is available in export\TS_AIMS_Torres Strait-Sentinel-2_Composite.tif.

Change Log: 2023-03-02: Eric Lawrey Created a merged version of the satellite imagery, with no alpha blending so that it can be used in ArcGIS Pro. It is now a single large GeoTiff image. The Google Earth Engine source code for the World_AIMS_Marine-satellite-imagery was included to improve the reproducibility and provenance of the dataset, along with a calculation of the distribution of image dates that went into the final composite image. A WMS service for the imagery was also setup and linked to from the metadata. A cross reference to the older Torres Strait clear sky clear water Landsat composite imagery was also added to the record.

22 Nov 2023: Eric Lawrey Added the data and maps for close up of Mer. - 01-data/TS_DNRM_Mer-aerial-imagery/ - preview/Torres-Strait-Mer-Map-Landscape-A0.jpeg - exports/Torres-Strait-Mer-Map-Landscape-A0.pdf Updated 02-Torres-Strait-regional-maps.qgz to include the layout for the new map.

Source datasets: Complete Great Barrier Reef (GBR) Island and Reef Feature boundaries including Torres Strait Version 1b (NESP TWQ 3.13, AIMS, TSRA, GBRMPA), https://eatlas.org.au/data/uuid/d2396b2c-68d4-4f4b-aab0-52f7bc4a81f5

Geoscience Australia (2014b), Seas and Submerged Lands Act 1973 - Australian Maritime Boundaries 2014a - Geodatabase [Dataset]. Canberra, Australia: Author. https://creativecommons.org/licenses/by/4.0/ [license]. Sourced on 12 July 2017, https://dx.doi.org/10.4225/25/5539DFE87D895

Basemap/AU_GA_AMB_2014a/Exclusive_Economic_Zone_AMB2014a_Limit.shp The original data was obtained from GA (Geoscience Australia, 2014a). The Geodatabase was loaded in ArcMap. The Exclusive_Economic_Zone_AMB2014a_Limit layer was loaded and exported as a shapefile. Since this file was small no clipping was applied to the data.

Geoscience Australia (2014a), Treaties - Australian Maritime Boundaries (AMB) 2014a [Dataset]. Canberra, Australia: Author. https://creativecommons.org/licenses/by/4.0/ [license]. Sourced on 12 July 2017, http://dx.doi.org/10.4225/25/5539E01878302 Basemap/AU_GA_Treaties-AMB_2014a/Papua_New_Guinea_TSPZ_AMB2014a_Limit.shp The original data was obtained from GA (Geoscience Australia, 2014b). The Geodatabase was loaded in ArcMap. The Papua_New_Guinea_TSPZ_AMB2014a_Limit layer was loaded and exported as a shapefile. Since this file was small no clipping was applied to the data.

AIMS Coral Sea Features (2022) - DRAFT This is a draft version of this dataset. The region for Ashmore and Boot reef was checked. The attributes in these datasets haven't been cleaned up. Note these files should not be considered finalised and are only suitable for maps around Ashmore Reef. Please source an updated version of this dataset for any other purpose. CS_AIMS_Coral-Sea-Features/CS_Names/Names.shp CS_AIMS_Coral-Sea-Features/CS_Platform_adj/CS_Platform.shp CS_AIMS_Coral-Sea-Features/CS_Reef_Boundaries_adj/CS_Reef_Boundaries.shp CS_AIMS_Coral-Sea-Features/CS_Depth/CS_AIMS_Coral-Sea-Features_Img_S2_R1_Depth5m_Coral-Sea.shp CS_AIMS_Coral-Sea-Features/CS_Depth/CS_AIMS_Coral-Sea-Features_Img_S2_R1_Depth10m_Coral-Sea.shp

Murray Island 20 Sept 2011 15cm SISP aerial imagery, Queensland Spatial Imagery Services Program, Department of Resources, Queensland This is the high resolution imagery used to create the map of Mer.

Marine satellite imagery (Sentinel 2 and Landsat 8) (AIMS), https://eatlas.org.au/data/uuid/5d67aa4d-a983-45d0-8cc1-187596fa9c0c - World_AIMS_Marine-satellite-imagery

Data Location: This dataset is filed in the eAtlas enduring data repository at: data\custodian\2020-2029-AIMS\TS_AIMS_Torres-Strait-Sentinel-2-regional-maps. On the eAtlas server it is stored at eAtlas GeoServer\data\2020-2029-AIMS.

Why?

Object tracking, or more precisely the re-identification of objects in video streams, relies more and more on deep convolutional and residual networks. And they require a lot of good training data. Moreover, we want to show that including object mask in the alpha channel may pose additional benefits in object re-identification.

What?

The dataset was constructed by crunching image sequences from the Multiple Object Tracking Challenge 2016/7 dataset (they differ only in provided detections and ground truth, neither of which is used here). As a bonus, I have taken a random YouTube video in high resolution with people walking around (youtu.be/NEfxRHeb-70) and extracted five tracklets from there. Mask R-CNN provides a proposal of object mask which is stored in the alpha channel in

Files are organised similarly as in the MARS dataset, one of the most prevalent in object re-identification learning. Just a couple of notes here: - Images are actually in four-channel PNG (RGBA) with aspect ratio 1:2, object centred in the bounding box, padded with zeros. - Opposed to MARS, each tracklet is considered a new sequence. This may be suboptimal as the same person can be in multiple tracklets. - The train/test split is approx. 50:50, IDs do not overlap.

This dataset contains Landsat 5 imagery for selected areas of Queensland, currently Torres Strait and around Lizard Island and Cape Tribulation.

This collection was made as a result of the development of the Torres Strait Features dataset. It includes a number (typically 4 - 8) of selected Landsat images for each scene from the entire Landsat 5 archive. These images were selected for having low cloud cover and clear water. The aim of this collection was to allow investigation of the marine features.

The complete catalogue of Landsat 5 for scenes 96_70, 96_71, 97_67, 97_68, 98_66, 98_67, 98_68_99_66, 99_67 and 99_68 were downloaded from the Google Earth Engine site ( https://console.developers.google.com/storage/earthengine-public/landsat/ ). The images were then processed into low resolution true colour using GDAL. They were then reviewed for picture clarity and the best ones were selected and processed at full resolution to be part of this collection.

The true colour conversion was achieved by applying level adjustment to each channel to ensure that the tonal scaling of each channel was adjusted to give a good overall colour balance. This effectively set the black point of the channel and the gain. This adjustment was applied consistently to all images.

• Red: Channel B3, Black level 8, White level 58
• Green: Channel B2, Black level 10, White level 55
• Blue: Channel B1, Black level 32, White level 121

Note: A constant level adjustment was made to the images regardless of the time of the year that the images were taken. As a result images in the summer tend to be brighter than those in the winter.

After level adjustment the three channels were merged into a single colour image using gdal_merge. The black surround on the image was then made transparent using the GDAL nearblack command.

This collection consists of 59 images saved as 4 channel (Red, Green, Blue, Alpha) GeoTiff images with LZW compression (lossless) and internal overviews with a WGS 84 UTM 54N projection.

Each of the individual images can be downloaded from the eAtlas map client (Overlay layers: eAtlas/Imagery Base Maps Earth Cover/Landsat 5) or as a collection of all images for each scene.

Data Location:

This dataset is filed in the eAtlas enduring data repository at: data\NERP-TE\13.1_eAtlas\QLD_NERP-TE-13-1_eAtlas_Landsat-5_1988-2011

Training and test datasets for the deep learning model created in: A Deep Learning Approach to Identify and Segment α-Smooth Muscle Actin Stress Fiber Positive Cells. Contains 300, 3 channel training images (red: F-actin, green: a-SMA, blue: DAPI), rgb composite images, and segmentation labels (label 3 classes: background, a-SMA SF-, and a-SMA SF+)

4 channel images of cells used in the manuscript channel 1: F-Actin channel 2: alpha smooth muscle actin channel 3: DAPI channel 4: Phase-contrast

A code repository using this dataset and some sample generation results is at https://github.com/rubyyhj/PokeTypeGAN.

Content

This dataset contains in total 10,437 images (half of them are shiny variants) in 96x96 resolution from 898 Pokemon in different games, and their corresponding labels that may relate to their design. Labels include: - type1: str - type2: str - primary_color: str - legendary: bool - mega_evolution: bool - alolan_form: bool - galarian_form: bool - gigantamax: bool

Below are the collected images of Bulbaraur:

Gen3 E	Gen3 E (Frame 2)	Gen3 FL	Gen4 DP
https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44084dzj202o02ojr5.jpg" alt="1-gen3_e">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s443w5bhj202o02oq2p.jpg" alt="1-gen3_e">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s448n6rnj202o02oq2p.jpg" alt="1-gen3_e-frame2">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s449yt3qj202o02ot8h.jpg" alt="1-gen3_e-frame2">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44ddqeij202o02ojr5.jpg" alt="1-gen3_fl">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44ivp4hj202o02ot8h.jpg" alt="1-gen3_fl">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44n23xsj202o02owe9.jpg" alt="1-gen4_dp">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44rt84xj202o02o0si.jpg" alt="1-gen4_dp">

Gen4 HS	Gen4 HS (Frame 2)	Gen5	Gen5 (Back)
https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44uk1ccj202o02ot8h.jpg" alt="1-gen4_hs">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s44ypz9tj202o02ot8h.jpg" alt="1-gen4_hs">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s4511vlyj202o02ot8h.jpg" alt="1-gen4_hs-frame2">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s455gbezj202o02ot8h.jpg" alt="1-gen4_hs-frame2">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s456rn88j202o02oq2p.jpg" alt="1-gen5">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s45930ibj202o02oq2p.jpg" alt="1-gen5">	https://tva1.sinaimg.cn/large/e6c9d24egy1h1s45b1szzj202o02ogld.jpg" alt="1-gen5">https://tva1.sinaimg.cn/large/e6c9d24egy1h1s45e9ffqj202o02ogld.jpg" alt="1-gen5">

Data collection

This dataset is collected from public-available data and arranged according to GAN training needs. The metadata (Pokedex) is extracted and modified from: https://www.kaggle.com/datasets/kylekohnen/complete-pokemon-data-set
The raw sprites of Gen3, 4, 5 is downloaded from https://veekun.com/dex/downloads The raw sprites of Gen6, 7, 8 is downloaded from https://www.smogon.com/forums/threads/x-y-sprite-project.3486712/, https://www.smogon.com/forums/threads/sun-moon-sprite-project.3577711/ and https://www.smogon.com/forums/threads/smogon-sprite-project.3647722/, respectively. Only use “Back” sprites of Gen5 and after. Because before Gen5, the Back sprites are clipped instead of showing the full body of Pokemon.

Preprocessing

The alpha channel of the raw .png files is removed and the background is set to be white. Raw sprites in resolution lower than 96x96 (Gen3 and Gen4 sprites) are padded with white pixels to 96x96.