Context PGD MNIST is a dataset of adversarial images consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. PGD MNIST intends to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms on adversarial examples. It shares the same image size and structure of training and testing splits.

The original MNIST dataset contains a lot of handwritten digits. Members of the AI/ML/Data Science community love this dataset and use it as a benchmark to validate their algorithms. In fact, MNIST is often the first dataset researchers try. "If it doesn't work on MNIST, it won't work at all", they said. "Well, if it does work on MNIST, it may still fail on others."

Check your models accuracy on this dataset and improve the adversarial robustness of the models

Content Each image is 28 pixels in height and 28 pixels in width, for a total of 784 pixels in total. Each pixel has a single pixel-value associated with it, indicating the lightness or darkness of that pixel, with higher numbers meaning darker. This pixel-value is an integer between 0 and 255. The training and test data sets have 785 columns. The first column consists of the class labels (see above), and represents an integer. The rest of the columns contain the pixel-values of the associated image.

To locate a pixel on the image, suppose that we have decomposed x as x = i * 28 + j, where i and j are integers between 0 and 27. The pixel is located on row i and column j of a 28 x 28 matrix. For example, 31 indicates the pixel that is in the fourth column from the left, and the second row from the top, as in the ascii-diagram below.

Columns description Each row is a separate image Column 1 is the class label (integers 0-9). Remaining columns are pixel numbers (784 total). Each value is the darkness of the pixel (1 to 255)

Summary

This dataset contains corrected pixel anomalies from the CryoET dataset provided by @brendanartley.

Note: This dataset only includes the corrected samples.

Details of the Anomaly

Some of the original data (from the CryoET Data Portal) appear to have been incorrectly quantized, resulting in pixel values ranging from -128 to 127 — as if they were stored as int8, even though the expected data range should be from 0 to 255 like uint8. Which causes significant corruption of source image (see the images in [2]).

Correction Applied

The key correction applied is as follows:

x = x.astype("uint8").astype("float32")

Additional Information

• The voxel size is the same as in @brendanartley's dataset: (128, 512, 512).

References

• [1] The original dataset: https://www.kaggle.com/datasets/brendanartley/cryoet-flagellar-motors-dataset
• [2] https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/575028#3187364

The INDIGO Change Detection Reference Dataset

Description

This graffiti-centred change detection dataset was developed in the context of INDIGO, a research project focusing on the documentation, analysis and dissemination of graffiti along Vienna's Donaukanal. The dataset aims to support the development and assessment of change detection algorithms.

The dataset was collected from a test site approximately 50 meters in length along Vienna's Donaukanal during 11 days between 2022/10/21 and 2022/12/01. Various cameras with different settings were used, resulting in a total of 29 data collection sessions or "epochs" (see "EpochIDs.jpg" for details). Each epoch contains 17 images generated from 29 distinct 3D models with different textures. In total, the dataset comprises 6,902 unique image pairs, along with corresponding reference change maps. Additionally, exclusion masks are provided to ignore parts of the scene that might be irrelevant, such as the background.

To summarise, the dataset, labelled as "Data.zip," includes the following:

• Synthetic Images: These are colour images created within Agisoft Metashape Professional 1.8.4, generated by rendering views from 17 artificial cameras observing 29 differently textured versions of the same 3D surface model.
• Change Maps: Binary images that were manually and programmatically generated, using a Python script, from two synthetic graffiti images. These maps highlight the areas where changes have occurred.
• Exclusion Masks: Binary images are manually created from synthetic graffiti images to identify "no data" areas or irrelevant ground pixels.

Image Acquisition

Image acquisition involved the use of two different camera setups. The first two datasets (ID 1 and 2; cf. "EpochIDs.jpg") were obtained using a Nikon Z 7II camera with a pixel count of 45.4 MP, paired with a Nikon NIKKOR Z 20 mm lens. For the remaining image datasets (ID 3-29), a triple GoPro setup was employed. This triple setup featured three GoPro cameras, comprising two GoPro HERO 10 cameras and one GoPro HERO 11, all securely mounted within a frame. This triple-camera setup was utilised on nine different days with varying camera settings, resulting in the acquisition of 27 image datasets in total (nine days with three datasets each).

Data Structure

The "Data.zip" file contains two subfolders:

• 1_ImagesAndChangeMaps: This folder contains the primary dataset. Each subfolder corresponds to a specific epoch. Within each epoch folder resides a subfolder for every other epoch with which a distinct epoch pair can be created. It is important to note that the pairs "Epoch Y and Epoch Z" are equivalent to "Epoch Z and Epoch Y", so the latter combinations are not included in this dataset. Each sub-subfolder, organised by epoch, contains 17 more subfolders, which hold the image data. These subfolders consist of:
  • Two synthetic images rendered from the same synthetic camera ("X_Y.jpg" and "X_Z.jpg")
  • The corresponding binary reference change map depicting the graffiti-related differences between the two images ("X_YZ.png"). Black areas denote new graffiti (i.e. "change"), and white denotes "no change". "DataStructure.png" provides a visual explanation concerning the creation of the dataset.
    
    The filenames follow the following pattern:
    • X - Is the ID number of the synthetic camera. In total, 17 synthetic cameras were placed along the test site
    • Y - Corresponds to the reference epoch (i.e. the "older epoch")
    • Z - Corresponds to the "new epoch"
      
• 2_ExclusionMasks: This folder contains the binary exclusion masks. They were manually created from synthetic graffiti images and identify "no data" areas or areas considered irrelevant, such as "ground pixels". Two exclusion masks were generated for each of the 17 synthetic cameras:
  • "groundMasks": depict ground pixels which are usually irrelevant for the detection of graffiti
  • "noDataMasks": depict "background" for which no data is available.
    

A detailed dataset description (including detailed explanations of the data creation) is part of a journal paper currently in preparation. The paper will be linked here for further clarification as soon as it is available.

Licensing

Due to the nature of the three image types, this dataset comes with two licenses:

• Synthetic images:
  • These come with an In Copyright license (for the rights usage terms, see https://rightsstatements.org/page/InC/1.0/?language=en).
  • The copyright lies with:
    • the Ludwig Boltzmann Gesellschaft (https://d-nb.info/gnd/1024204324)
    • the TU Wien (https://d-nb.info/gnd/55426-1)
    • One or more anonymous graffiti creator(s) upon whose work these images are based.
  • The first two entities are also the licensor of these images.
    
• Change maps and masks:
  • These are openly licensed via CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0)
  • In this case, the copyright lies with:
    • the Ludwig Boltzmann Gesellschaft (https://d-nb.info/gnd/1024204324)
    • the TU Wien (https://d-nb.info/gnd/55426-1)
  • Both institutes are also the licensor of these images.

Every synthetic image, change map and mask has this licensing information embedded as IPTC photo metadata. In addition, the images' IPTC metadata also provide a short image description, the image creator and the creator's identity (in the form of an ORCiD).

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

If there are any questions, problems or suggestions for the dataset or the description, please do not hesitate to contact the corresponding author, Benjamin Wild.

This data set provides hot pixel data, as an indicator of fires that were detected by the GOES-8 satellite for the state of Acre, Brazil. Image data were collected for extended periods over the course of 3 years (1998, 2000 and 2001). Data were filtered to select only pixels identified and processed by the GOES-8 Automated Biomass Burning Algorithm (ABBA), where estimates of sub-pixel fire characteristics including size and temperature were able to be determined. There are three comma-delimited ASCII data files with this data set.

From Pixels to Prose: A Large Dataset of Dense Image Captions

[ arXiv paper ] PixelProse is a comprehensive dataset of over 16M (million) synthetically generated captions, leveraging cutting-edge vision-language models (Gemini 1.0 Pro Vision) for detailed and accurate descriptions.

  1. Details

Total number of image-caption pairs: 16,896,214 (16.9M)

6,538,898 (6.5M) pairs in the split of CommonPool 9,066,455 (9.1M) pairs in the split of CC12M 1,290,861 (1.3M) pairs in… See the full description on the dataset page: https://huggingface.co/datasets/tomg-group-umd/pixelprose.

Auxiliary information for pixel cloud product indicating to which water bodies the pixels are assigned in river and lake products. Also includes height-constrained pixel geolocation after reach- or lake-scale averaging. Point cloud over tile (approx 64x64 km2); half swath (left or right side of full swath). Available in netCDF-4 file format.

WELDAKLL.015 was decommissioned on December 2, 2019. Users are encouraged to use the improved monthly Global Web-Enabled Landsat Data (GWELD) Version 3, 3.1, and 3.2 datasets.The NASA Web-Enabled Landsat Data (WELD) Alaska Latitude Longitude (LL) products are defined with 30 meter (m) composite pixel of the center latitude/longitude point. The NASA WELD data contain a range of values for the minimum/maximum latitude and longitude points. The LL products were created to account for the actual center pixel latitude and longitude points.WELD systematically generated 30 m composited Landsat Enhanced Thematic Mapper Plus (ETM+) mosaics of the United States and Alaska from 2002 to 2012 to provide consistent data for deriving land cover as well as geophysical and biophysical products for regional assessments of surface dynamics for effective study of Earth system function. WELDAKLL is distributed in Hierarchical Data Format 4 (HDF4).The WELD project is funded by the National Aeronautics and Space Administration (NASA) and is a collaboration between the United States Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center and the South Dakota State University (SDSU) Geospatial Sciences Center of Excellence (GSCE). Known Issues WELD Version 1.5 known issues can be found in the WELD Version 1.5 User Guide.Improvements/Changes from Previous Version Version 1.5 is the original version.

This dataset is derived from the USGS Burned Area Products (Hawbaker et al. 2017). We used Burned Area (BA) version 2 products (USGS 2019). We evaluated the annualBAECV Burn Probability (BP) datasets –which are raster datasets – for evidence of burns. The annual datasets span an entire calendar year (e.g.,Jan 1 through Dec 31) and indicate the maximum BP within the year (0-100%). For each year between 1994 and 2020, we combined the annual datasets of interest within individual ARD Tiles into a single annual raster dataset (i.e., we mosaicked the tiles) for further processing. We performed all additional processing steps on the annual mosaicked datasets as this provided statewide consistency. We identified pixels as burned or unburned according to their probability value; initially, we retained all pixels with an annual BP between 85-100% based on Hawbaker et al. (2017). Values between 90-100% were then converted to presence/absence rasters and we used image processing methods to remove ‘speckling’ (e.g.,fill in small holes within a burned area and remove groups of pixels less than a specified size/amount). This process resulted in annual rasters and vectors indicating burn presence (with 90-100% probability) for groups of pixels greater than ~2.24 acres (e.g.,10 30m pixels, in any arrangement). We also assigned dates from the Burn Date (BD) dataset to these same pixels as a surrogate for seasonality. We evaluated these products against fire records for three pilot areas. For each area, we held a meeting with fire managers, either in person or via web conferencing methods. We invited managers to inspect the data with us to evaluate their thoughts on the products. Through this process, managers provided many explanations for why no burn was detected and where/why fire detection was performing very well, as well as some ideas and suggestions for moving forward (all of which we relayed to USGS). Many of these comments reflect known limitations previously documented (see Hawbaker et al. 2017, Vanderhoof et al. 2017). Based on these meetings, we have applied the processing “logic” across the entire state at 90-100%BP. Fire regime metrics such as number of times burned, year last burned, and time since previous fire (as measured from 2020) are included in the dataset upload, and were derived using these annual presence/absence rasters and vectors.

WELDUSLL.015 was decommissioned on December 2, 2019. Users are encouraged to use the improved monthly Global Web-Enabled Landsat Data (GWELD) Version 3, 3.1, and 3.2 datasets. NASA’s Web-Enabled Landsat Data (WELD) are generated from composited 30 meter (m) Landsat Enhanced Thematic Mapper Plus (ETM+) mosaics of the United States and Alaska from 2002 to 2012. These mosaics provide consistent data to derive land cover as well as geophysical and biophysical products for regional assessments of surface dynamics for effective study of Earth system function. The Conterminous (CONUS) United States Latitude/Longitude (WELDUSLL) products provide the geographic latitude and longitude for the center of each 30 meter (m) pixel within a tile of WELD data. WELDUSLL is distributed in Hierarchical Data Format 4 (HDF4).The WELD project is funded by the National Aeronautics and Space Administration (NASA) and is a collaboration between the United States Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center and the South Dakota State University (SDSU) Geospatial Sciences Center of Excellence (GSCE). Known Issues WELD Version 1.5 known issues can be found in the WELD Version 1.5 User Guide.Improvements/Changes from Previous Version Version 1.5 is the original version.

Introduction and Rationale:Due to our increasing understanding of the role the surrounding landscape plays in ecological processes, a detailed characterization of land cover, including both agricultural and natural habitats, is ever more important for both researchers and conservation practitioners. Unfortunately, in the United States, different types of land cover data are split across thematic datasets that emphasize agricultural or natural vegetation, but not both. To address this data gap and reduce duplicative efforts in geospatial processing, we merged two major datasets, the LANDFIRE National Vegetation Classification (NVC) and USDA-NASS Cropland Data Layer (CDL), to produce integrated ‘Spatial Products for Agriculture and Nature’ (SPAN). Our workflow leveraged strengths of the NVC and the CDL to produce detailed rasters comprising both agricultural and natural land-cover classes. We generated SPAN for each year from 2012-2021 for the conterminous United States, quantified agreement between input layers and accuracy of our merged product, and published the complete workflow necessary to update SPAN. In our validation analyses, we found that approximately 5.5% of NVC agricultural pixels conflicted with the CDL, but we resolved a majority of these conflicts based on surrounding agricultural land, leaving only 0.6% of agricultural pixels unresolved in the final version of SPAN.Contents:Spatial dataNational rasters of land cover in the conterminous United States: 2012-2021Rasters of pixels mismatched between CDL and NVC: 2012-2021Resources in this dataset:Resource Title: SPAN land cover in the conterminous United States: 2012-2021 - SCINet File Name: KammererNationalRasters.zip Resource Description: GeoTIFF rasters showing location of pixels that are mismatched between 2016 NVC and specific year of CDL (2012-2021). Spatial Products for Agriculture and Nature ('SPAN') land cover in the conterminous United States from 2012-2021. This raster dataset is available in GeoTIFF format and was created by joining agricultural classes from the USDA-NASS Cropland Data Layer (CDL) to national vegetation from the LANDFIRE National Vegetation Classification v2.0 ('Remap'). Pixels of national vegetation are the same in all rasters provided here and represent land cover in 2016. Agricultural pixels were taken from the CDL in the specified year, so depict agricultural land from 2012-2021. Resource Title: Rasters of pixels mismatched between CDL and NVC: 2012-2021 - SCINet File Name: MismatchedNational.zip Resource Description: GeoTIFF rasters showing location of pixels that are mismatched between 2016 NVC and specific year of CDL (2012-2021). This dataset includes pixels that were classified as agriculture in the NVC but, in the CDL, were not agriculture (or were a conflicting agricultural class). For more details, we refer users to the linked publication describing our geospatial processing and validation workflow.SCINet users: The files can be accessed/retrieved with valid SCINet account at this location: /LTS/ADCdatastorage/NAL/published/node455886/ See the SCINet File Transfer guide for more information on moving large files: https://scinet.usda.gov/guides/data/datatransferGlobus users: The files can also be accessed through Globus by following this data link. The user will need to log in to Globus in order to retrieve this data. User accounts are free of charge with several options for signing on. Instructions for creating an account are on the login page.

A comprehensive benchmark of the DECIMER Image Transformer was conducted using all publicly available OCSR benchmark datasets and DECIMER test datasets.

USPTO: A set of 5,719 images of chemical structures and the corresponding MOL files (US Patent Office) obtained from the OSRA online presence

UOB: The dataset of 5,740 images and MOL files of chemical structures developed by the University of Birmingham, United Kingdom, and published alongside MolRec

CLEF: The Conference and Labs of the Evaluation Forum test set of 992 images and molfiles published in 2012

JPO: A subset (450 images and MOL files) of a dataset based on data from the Japanese Patent Office, obtained from the OSRA online presence. Note that this dataset contains many labels (sometimes with Japanese characters) and irregular features, such as variations in the line thickness. Additionally, some images have poor quality and contain a lot of noise.

RanDepict250k: A set of 250,000 chemical structure depictions generated with RanDepict (1.0.8) using RanDepict’s depiction feature fingerprints to ensure diverse depiction parameters. None of the depicted molecules is present in the DECIMER training data. The images here are all 299 x 299 pixels in size.

RanDepict250k_augmented: A set of the same 250,000 images from the RanDepict250k dataset. Additional augmentations (examples: mild rotation, shearing, insertion of labels and reaction arrows around the structures, insertion of curved arrows in the structure) were added to the images using RanDepict. The images here are all 299 x 299 pixels in size.

DECIMER hand-drawn: A set of 5,088 chemical structure depictions which were manually drawn by a group of 24 volunteers. The drawn molecules have been picked using the MaxMin algorithm from all molecules in PubChem so that the set represents a big part of the chemical space.

Indigo: 50,000 images generated by Staker et al. using Indigowhich were collected from the supplementary information. All images have a resolution of 224 x 224 pixels.

USPTO_big: 50,000 images from the USPTO from Staker et al. which were collected from the supplementary information. All images have a resolution of 224 x 224 pixels.

Img2Mol test set: A set of 25,000 chemical structure depictions used by Clévert et al. for testing . All images have a resolution of 224 x 224 pixels.

The dataset consist of three categories; image subsets, burned area masks and quicklooks. The image subsets are derived from Landsat-8 scenes taken during the years 2019 and 2021. Each image has a size of 512x512 pixels and consists of 8 multispectral. The sequence of band names from band 1 to band 7 of the image subset is same as the sequence of band names of landsat-8 scene, except for band 8 of the image subset which is band 9 (cirrus band) in the original landsat-8 scene. The image subsets are saved in GeoTIFF file format with the latitude longitude coordinate system and WGS 1984 as the datum. The spatial resolution of image subsets is 0.00025 degree and the pixel values are stored in 16 bit unsigned integer with the range of value from 0 to 65535. The total of the dataset is 227 images which containing object of burned area surrounded by various ecological diversity backgrounds such as forest, shrub, grassland, waterbody, bare land, settlement, cloud and cloud shadow. In some cases, there are some image subsets with the burned areas covered by smoke due to the fire is still active. Some image subsets also overlap each other to cover the area of burned scar which the area is too large. The burned area mask is a binary annotation image which consists of two classes; burned area as the foreground and non-burned area as the background. These binary images are saved in 8 bit unsigned integer where the burned area is indicated by the pixel value of 1, whereas the non-burned area is indicated by 0. The burned area masks in this dataset contain only burned scars and are not contaminated with thick clouds, shadows, and vegetation. Among 227 images, 206 images contain burned areas whereas 21 images contain only background. The highest number of images in this dataset is dominated by images with coverage percentage of burned area between 0 and 10 percent. Our dataset also provides quicklook image as a quick preview of image subset. It offers a fast and full size preview of image subset without opening the file using any GIS software. The quicklook images can also be used for training and evaluating the model as a substitute of image subsets. The image size is 512x512 pixels same as the size of image subset and annotation image. It consists of three bands as a false color composite quicklook images, with combination of band 7 (SWIR-2), band 5 (NIR), and band 4 (red). These RGB composite images have been performed contrast stretching to enhance the images visualizations. The quicklook images are stored in GeoTIFF file format with 8 bit unsigned integer.

This work was financed by Riset Inovatif Produktif (RISPRO) fund through Prioritas Riset Nasional (PRN) project, grant no. 255/E1/PRN/2020 for 2020 - 2021 contract period.

A dataset suitable for spatiotemporal object detection is constructed using several aerial video clips of traffic in different road segments in Nicosia, Cyprus, captured using UAVs, rather than single areas in low resolution satelite images as other datasets. By compiling multiple sequences of images extracted from these videos, the dataset accumulates a substantial corpus of 6,600 frames. The dataset encapsulates 3 classes: ‘car’, ‘truck’ and ‘bus’ with a distribution of 81165, 1541, and 1625 respectively in the case that we only use the even frame annotations, which approximately doubles when considering the entire dataset. An additional challenge of the dataset that mirrors real world application is the fact that the classes are not balanced, as there is a significantly larger number of cars compared to trucks and buses, as in a regular transportation network. The images have Full-HD resolution, with object sizes approximately between 20x20 to 150x150 pixels. The dataset was prepared in the YOLO format. The dataset was split into 80% for training and the remaining 20% for validation. The importance of such a dataset lies in its capability to encapsulate both spatial and temporal nuances. We note the frames belonging in the same continuous sequence as such the dataset can potentially be used to develop approaches that operate on multiple sequential frames for object detection by sampling a number of frames from the same sequence. Dataset Feature Description Total Images ~6600 Image Sizes 1920x1080 Classes Car,Bus,Truck Data Collection Collect from UAVs at different locations in Nicosia, Cyprus Data Format PNG Labelling Format YOLO

The Road Scenes Panoptic Segmentation Dataset is aimed at applications in visual entertainment and autonomous driving, featuring a collection of internet-collected road scene images with resolutions exceeding 1600 x 1200 pixels. This dataset specializes in panoptic segmentation, annotating every identifiable instance within the images, such as vehicles, roads, lane lines, vegetation, and people, providing a detailed dataset for comprehensive road scene analysis.

The Blur Area Segmentation Dataset is designed for use in robotics and visual entertainment, composed of internet-collected images with resolutions ranging from 960 x 720 to 1024 x 768 pixels. This dataset focuses on semantic segmentation, specifically targeting blue areas within images. Each blue area is annotated at the pixel level, providing valuable data for applications requiring color-based segmentation or analysis.

The Harmonized Landsat Sentinel-2 (HLS) project provides consistent surface reflectance data from the Operational Land Imager (OLI) aboard the joint NASA/USGS Landsat 8 satellite and the Multi-Spectral Instrument (MSI) aboard Europe’s Copernicus Sentinel-2A, Sentinel-2B, and Sentinel-2C satellites. The combined measurement enables global observations of the land every 2–3 days at 30-meter (m) spatial resolution. The HLS project uses a set of algorithms to obtain seamless products from OLI and MSI that include atmospheric correction, cloud and cloud-shadow masking, spatial co-registration and common gridding, illumination and view angle normalization, and spectral bandpass adjustment. The HLSS30 product provides 30-m Nadir Bidirectional Reflectance Distribution Function (BRDF)-Adjusted Reflectance (NBAR) and is derived from Sentinel-2A, Sentinel-2B, and Sentinel-2C MSI data products. The HLSS30 and HLSL30 products are gridded to the same resolution and Military Grid Reference System (MGRS) tiling system and thus are “stackable” for time series analysis.The HLSS30 product is provided in Cloud Optimized GeoTIFF (COG) format, and each band is distributed as a separate COG. There are 13 bands included in the HLSS30 product along with four angle bands and a quality assessment (QA) band. See the User Guide for a more detailed description of the individual bands provided in the HLSS30 product.Known Issues Unrealistically high aerosol and low surface reflectance over bright areas: The atmospheric correction over bright targets occasionally retrieves unrealistically high aerosol and thus makes the surface reflectance too low. High aerosol retrievals, both false high aerosol and realistically high aerosol, are masked when quality bits 6 and 7 are both set to 1 (see Table 9 in the User Guide); the corresponding spectral data should be discarded from analysis. Issues over high latitudes: For scenes greater than or equal to 80 degrees north, multiple overpasses can be gridded into a single MGRS tile resulting in an L30 granule with data sensed at two different times. In this same area, it is also possible that Landsat overpasses that should be gridded into a single MGRS tile are actually written as separate data files. Finally, for scenes with a latitude greater than or equal to 65 degrees north, ascending Landsat scenes may have a slightly higher error in the BRDF correction because the algorithm is calibrated using descending scenes. Fmask omission errors: There are known issues regarding the Fmask band of this data product that impacts HLSL30 data prior to April of 2022. The HLS Fmask data band may have omission errors in water detection for cases where water detection using spectral data alone is difficult, and omission and commission errors in cloud shadow detection for areas with great topographic relief. This issue does not impact other bands in the dataset. Inconsistent snow surface reflectance between Landsat and Sentinel-2: The HLS snow surface reflectance can be highly inconsistent between Landsat and Sentinel-2. When assessed on same-day acquisitions from Landsat and Sentinel-2, Landsat reflectance is generally higher than Sentinel-2 reflectance in the visible bands. Unrealistically high snow surface reflectance in the visible bands: By design, the Land Surface Reflectance Code (LaSRC) atmospheric correction does not attempt aerosol retrieval over snow; instead, a default aerosol optical thickness (AOT) is used to drive the snow surface reflectance. If the snow detection fails, the full LaSRC is used in both AOT retrieval and surface reflectance derivation over snow, which produces surface reflectance values as high as 1.6 in the visible bands. This is a common problem for spring images at high latitudes. Unrealistically low surface reflectance surrounding snow/ice: Related to the above, the AOT retrieval over snow/ice is generally too high. When this artificially high AOT is used to derive the surface reflectance of the neighboring non-snow pixels, very low surface reflectance will result. These pixels will appear very dark in the visible bands. If the surface reflectance value of a pixel is below -0.2, a NO_DATA value of -9999 is used. In Figure 1, the pixels in front of the glaciers have surface reflectance values that are too low. * Unrealistically low reflectance surrounding clouds: Like for snow, the HLS atmospheric correction does not attempt aerosol retrieval over clouds and a default AOT is used instead. But if the cloud detection fails, an artificially high AOT will be retrieved over clouds. If the high AOT is used to derive the surface reflectance of the neighboring cloud-free pixels, very low surface reflectance values will result. If the surface reflectance value of a pixel is below -0.2, a NO_DATA value of -9999 is used. * Unusually low reflectance around other bright land targets: While the HLS atmospheric correction retrieves AOT over non-cloud, non-snow bright pixels, the retrieved AOT over bright targets can be unrealistically high in some cases, similar to cloud or snow. If this unrealistically high AOT is used to derive the surface reflectance of the neighboring pixels, very low surface reflectance values can result as shown in Figure 2. If the surface reflectance value of a pixel is below -0.2, a NO_DATA value of -9999 is used. These types of bright targets are mostly man-made, such as buildings, parking lots, and roads. * Dark plumes over water: The HLS atmospheric correction does not attempt aerosol retrieval over water. For water pixels, the AOT retrieved from the nearest land pixels is used to derive the surface reflectance, but if the retrieval is incorrect, e.g. from a cloud pixel, this high AOT will create dark stripes over water, as shown in Figure 3. This happens more often over large water bodies, such as lakes and bays, than over narrow rivers. * Landsat WRS-2 Path/Row boundary in L30 reflectance: HLS performs atmospheric correction on Landsat Level 1 images in the original Worldwide Reference System 2 (WRS2) path/row before the derived surface reflectance is reprojected into Military Grid Reference System (MGRS) tiles. If a WRS-2 Landsat image is very cloudy, the AOT from a few remaining clear pixels might be used for the atmospheric correction of the entire image. The AOT that is used can be quite different from the value for the adjacent row in the same path, which results in an artificial abrupt change from one row to the next, as shown in Figure 4. This occurrence is very rare. * Landsat WRS2 path/row boundary in cloud masks: The cloud mask algorithm Fmask creates mask labels by applying thresholds to the histograms of some metrics for each path/row independently. If two adjacent rows in the same path have distinct distributions within the metrics, abrupt changes in masking patterns can appear across the row boundary, as shown in Figure 5. This occurrence is very rare. * Fmask configuration was deficient for 2-3 months in 2021: The HLS installation of Fmask failed to include auxiliary digital elevation model (DEM) and European Space Agency (ESA) Global Surface Water Occurrence data for a 2-3 month run in 2021. This impacted the masking results over water and in mountainous regions. * The reflectance “scale_factor” and “offset” for some L30 and S30 bands were not set: The HLS reflectance scaling factor is 0.0001 and offset is 0. However, this information was not set in the Cloud Optimized GeoTIFF (COG) files of some bands for a small number of granules. The lack of this information creates a problem for automatic conversion of the reflectance data, requiring explicit scaling in applications. The problem has been corrected, but the affected granules have not been reprocessed. * Incomplete map projection information: For a time, HLS imagery was produced with an incomplete coordinate reference system (CRS). The metadata contains the Universal Transverse Mercator (UTM) zone and coordinates necessary to geolocate pixels within the image but might not be in a standard form, especially for granules produced early in the HLS mission. As a result, an error will occur in certain image processing packages due to the incomplete CRS. The simplest solution is to update to the latest version of Geospatial Data Abstraction Library (GDAL) and/or rasterio, which use the available information without error. * False northing of 10^7 for the L30 angle data: The L30 and S30 products do not use a false northing for the UTM projection, and the angle data are supposed to follow the same convention. However, the L30 angle data incorrectly uses a false northing of 10^7. There is no problem with the angle data itself, but the false northing needs to be set to 0 for it to be aligned with the reflectance. L30 from Landsat L1GT scenes: Landsat L1GT scenes were not intended for HLS due to their poor geolocation. However, some scenes made it through screening for a short period of HLS production. L1GT L30 scenes mainly consist of extensive cloud or snow that can be eliminated using the Fmask quality bits layer. Users can also identify an L1GT-originated L30 granule by examining the HLS cmr.xml metadata file. The UTC dates in the L30/S30 filenames may not be the local dates: UTC dates are used by ESA and the U.S. Geological Survey (USGS) in naming their Level 1 images, and HLS processing retains this information to name the L30 and S30 products. Landsat and Sentinel-2 overpass eastern Australia and New Zealand around 10AM local solar time, but this area is in either UTC+10:00 or +11:00 zone; therefore, the UTC time for some orbits is in fact near the end of the preceding UTC day. For example, HLS.S30.T59HQS.2016117T221552.v2.0 was acquired in the 22nd hour of day 117 of year 2016 in UTC, but the time was 10:15:52 of day 118 locally. Approximately 100 minutes later HLS.S30.T56JML.2016117T235252.v2.0 was acquired in the next orbit in eastern Australia. This issue also occurs for Landsat. For

The Head and Neck Semantic Segmentation Dataset is designed for the e-commerce & retail and media & entertainment sectors, featuring a collection of AI-generated cartoon images with resolutions above 1024 x 1024 pixels. This dataset focuses on semantic segmentation, specifically targeting the main character's head, including face, hair, and any accessories, as well as the neck area up to the collarbone, with an allowance for small, unsegmented parts on the edges.

Today, deep neural networks are widely used in many computer vision problems, also for geographic information systems (GIS) data. This type of data is commonly used for urban analyzes and spatial planning. We used orthophotographic images of two residential districts from Kielce, Poland for research including urban sprawl automatic analysis with Transformer-based neural network application.Orthophotomaps were obtained from Kielce GIS portal. Then, the map was manually masked into building and building surroundings classes. Finally, the ortophotomap and corresponding classification mask were simultaneously divided into small tiles. This approach is common in image data preprocessing for machine learning algorithms learning phase. Data contains two original orthophotomaps from Wietrznia and Pod Telegrafem residential districts with corresponding masks and also their tiled version, ready to provide as a training data for machine learning models.Transformed-based neural network has undergone a training process on the Wietrznia dataset, targeted for semantic segmentation of the tiles into buildings and surroundings classes. After that, inference of the models was used to test model's generalization ability on the Pod Telegrafem dataset. The efficiency of the model was satisfying, so it can be used in automatic semantic building segmentation. Then, the process of dividing the images can be reversed and complete classification mask retrieved. This mask can be used for area of the buildings calculations and urban sprawl monitoring, if the research would be repeated for GIS data from wider time horizon.Since the dataset was collected from Kielce GIS portal, as the part of the Polish Main Office of Geodesy and Cartography data resource, it may be used only for non-profit and non-commertial purposes, in private or scientific applications, under the law "Ustawa z dnia 4 lutego 1994 r. o prawie autorskim i prawach pokrewnych (Dz.U. z 2006 r. nr 90 poz 631 z późn. zm.)". There are no other legal or ethical considerations in reuse potential.Data information is presented below.wietrznia_2019.jpg - orthophotomap of Wietrznia districtmodel's - used for training, as an explanatory imagewietrznia_2019.png - classification mask of Wietrznia district - used for model's training, as a target imagewietrznia_2019_validation.jpg - one image from Wietrznia district - used for model's validation during training phasepod_telegrafem_2019.jpg - orthophotomap of Pod Telegrafem district - used for model's evaluation after training phasewietrznia_2019 - folder with wietrznia_2019.jpg (image) and wietrznia_2019.png (annotation) images, divided into 810 tiles (512 x 512 pixels each), tiles with no information were manually removed, so the training data would contain only informative tilestiles presented - used for the model during training (images and annotations for fitting the model to the data)wietrznia_2019_vaidation - folder with wietrznia_2019_validation.jpg image divided into 16 tiles (256 x 256 pixels each) - tiles were presented to the model during training (images for validation model's efficiency); it was not the part of the training datapod_telegrafem_2019 - folder with pod_telegrafem.jpg image divided into 196 tiles (256 x 265 pixels each) - tiles were presented to the model during inference (images for evaluation model's robustness)Dataset was created as described below.Firstly, the orthophotomaps were collected from Kielce Geoportal (https://gis.kielce.eu). Kielce Geoportal offers a .pst recent map from April 2019. It is an orthophotomap with a resolution of 5 x 5 pixels, constructed from a plane flight at 700 meters over ground height, taken with a camera for vertical photos. Downloading was done by WMS in open-source QGIS software (https://www.qgis.org), as a 1:500 scale map, then converted to a 1200 dpi PNG image.Secondly, the map from Wietrznia residential district was manually labelled, also in QGIS, in the same scope, as the orthophotomap. Annotation based on land cover map information was also obtained from Kielce Geoportal. There are two classes - residential building and surrounding. Second map, from Pod Telegrafem district was not annotated, since it was used in the testing phase and imitates situation, where there is no annotation for the new data presented to the model.Next, the images was converted to an RGB JPG images, and the annotation map was converted to 8-bit GRAY PNG image.Finally, Wietrznia data files were tiled to 512 x 512 pixels tiles, in Python PIL library. Tiles with no information or a relatively small amount of information (only white background or mostly white background) were manually removed. So, from the 29113 x 15938 pixels orthophotomap, only 810 tiles with corresponding annotations were left, ready to train the machine learning model for the semantic segmentation task. Pod Telegrafem orthophotomap was tiled with no manual removing, so from the 7168 x 7168 pixels ortophotomap were created 197 tiles with 256 x 256 pixels resolution. There was also image of one residential building, used for model's validation during training phase, it was not the part of the training data, but was a part of Wietrznia residential area. It was 2048 x 2048 pixel ortophotomap, tiled to 16 tiles 256 x 265 pixels each.

The Facial 17 Parts Segmentation Dataset is specifically compiled for the visual entertainment industry, featuring a range of internet-collected facial images with resolutions exceeding 1024 x 682 pixels. This dataset is dedicated to semantic segmentation, delineating 17 facial categories such as eyebrows, lips, eye pupils, and more. It also includes a selection of portrait images with occlusions, adding complexity and diversity to the dataset for more realistic application scenarios.

The Road Scene Semantic Segmentation Dataset is specifically designed for autonomous driving applications, featuring a collection of internet-collected images with a standard resolution of 1920 x 1080 pixels. This dataset is focused on semantic segmentation, aiming to accurately segment various elements of road scenes such as the sky, buildings, lane lines, pedestrians, and more, to support the development of advanced driver-assistance systems (ADAS) and autonomous vehicle technologies.