Find research datasets worth reusing

opencc-by-4.0May 2024View details →

Chang'e 5 Landing Camera Crater Detection Dataset

132 hand-labelled images from the Chang'e 5 Landing Camera. Visible impact craters in each image have their crater rim inscribed by a bounding ellipse. On average, there are approximately 50 labelled craters per image. The first 100 images of the landers descent were labelled - this is the intended training set. Every 10 images of the remaining 313 were then labelled - this is the intended testing set.   File Descriptions: CE5-ellipse-labels: joblib dump of ellipse parameters per image. change5-*.json: Raw labels as produced by the labelling software of choice, Label Studio.   Images: The images used in this work were produced and processed by the Ground Research and Application System (GRAS) of China's Lunar and Planetary Exploration Program (https://moon.bao.ac.cn). Specifically, the first 413 images from the Chang'e 5 landing camera level 2A were used. The images can be downloaded from here: <a href="https://moon.bao.ac.cn/ce5web/searchOrder_hyperSearchData.search?pid=CE5/LCAM/level/2A" target="_blank" rel="noopener">https://dx.doi.org/10.12350/CLPDS.GRAS.CE5.LCAM-2A.vA</a>.   Reference and Acknowledgement: Users of these annotations and associated data are requested to cite both the original dataset source (https://moon.bao.ac.cn) and the following paper: Matthew Rodda, Sofia McLeod, Ky Cuong Pham, and Tat-Jun Chin. (2024). Camera-Pose Robust Crater Detection from Chang'e 5. doi: https://doi.org/10.48550/arXiv.2406.04569   BibTeX: <pre><code>@misc{rodda2024camerapose, title={Camera-Pose Robust Crater Detection from Chang'e 5}, author={Matthew Rodda and Sofia McLeod and Ky Cuong Pham and Tat-Jun Chin}, year={2024}, eprint={2406.04569}, archivePrefix={arXiv}, primaryClass={cs.CV} }</code></pre>

FIGURE 8 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 8. Difference maps showing the percentage of events accurately detected by simulations without bioturbation subtracted from the percentage of events accurately detected by simulations with bioturbation. 100% sampling completeness and transition durations 0.001 times the event duration for A; 100% sampling completeness and transition durations five times the event duration for B; 25% sampling completeness and transition durations 0.001 times the event duration for C; 25% sampling completeness and transition durations five times the event duration for D. The solid black line marks the contour line for zero difference between the bioturbated and non-bioturbated simulations.

FIGURE 6 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 6. Effect of varying completeness with the duration of transition intervals. Simulation results with 25% completeness and transition lengths 0.001 times the event duration for A; and 25% completeness with transition lengths of five times the event duration for B. Background DCA-1 value is -0.5 and no bioturbation occurs. Excursion magnitude (y-axis) indicates the difference between the simulated background DCA-1 value and the simulated event DCA- 1 value. Resolution potential (x-axis) is the event duration divided by the time represented by the sample interval. Color shading indicates the median excursion magnitude of all samples that intersect an event. Contour lines show the parameter space where 50% (dashed line), 75% (dotted line) and 95% (solid line) of events are accurately detected.

FIGURE 12 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 12. Effect of sample thickness on accurate detection of events with excursion magnitudes of 2.8 DCA-1 units. Y-axis shows the percentage of accurately detected events at different sedimentation rates (x-axis) for three different event durations: 50 years, 100 years and 1000 years. Simulations are for the 2016 data set to approximate how a researcher might use a pilot data set to design a sampling procedure and are based on sampling with 25% completeness. A is simulations without the effect of bioturbation; B is simulations with the effect of bioturbation.

FIGURE 2 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 2. Simulation workflow in the paleontological assemblage mixer (paleoAM). A. Relative abundances of a given species (Epistominella pacifica) from the 355 samples of the 2021 data set, along the empirically derived DCA Axis 1 gradient. B. The per-bin mean of absolute abundance of E. pacifica across all samples within each bin. Absolute abundances are calculated from the relative abundances in A by rescaling the relative abundances to 10000 total specimens. C. Scaled kernel density estimates for E. pacifica, which depict the predicted abundance distribution of E. pacifica along DCA Axis 1 after fitting a kernel density estimate to the absolute abundances in B. D. Scaled kernel density estimates, like in C, for all taxa in the dataset showing their differing predicted abundance distributions along DCA Axis 1 with the kernel density of E. pacifica shown in C marked with an asterisk. E. Visual representation of parameters varied within the simulation along a vertical sediment core. From left to right: standard scenario, increased excursion magnitude, increased background value, increased resolution potential, sampling completeness, bioturbation and increased transition duration. Stacked rectangles represent potential sample intervals. In the first and last column, black dots indicate which intervals are sampled, gray dots indicate unsampled intervals. In the standard scenario, all potential sample intervals are sampled. Curved arrows denote sediment mixing among potential sample intervals.

FIGURE 1 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 1. Sample scores from detrended correspondence analysis (DCA) performed on benthic foraminiferal assemblages in the>63 µm size fraction from Integrated Ocean Drilling Program Expedition 341 Site U1419 in the Gulf of Alaska used in the simulation case study. A. DCA Axis 1 values for 355 assemblages from Sharon et al. (2021); B. DCA Axis 1 values for 47 assemblages representing a "pilot" data set of samples available and processed in 2016.

FIGURE 4 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 4. Percentage of events that are accurately detected and median excursion magnitudes for simulations performed with background DCA-1 values of -0.5, 0.5 and 1. For A and B, simulations used a background DCA-1 value of -0.5; for C and D, a background DCA-1 value of 0.5; for E and F, a background DCA-1 value of 1. Excursion magnitude (y-axis) indicates the difference between the simulated background DCA-1 value and the simulated event DCA- 1 value. Resolution potential (x-axis) is the event duration divided by the time represented by the sample interval. A, C and E. Color shading indicates the percentage of events that are accurately detected by at least one sample (i.e., the sample produces a DCA-1 value outside the 95% envelope of samples simulated at the background value, and within one DCA-1 unit of the simulated excursion magnitude). B, D and F. Color shading indicates the median excursion magnitude of all samples that intersect an event. Contour lines show the parameter space where 50% (dashed line), 75% (dotted line) and 95% (solid line) of events are accurately detected.

FIGURE 5 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 5. Effect of varying transition duration. Each panel represents a set of simulations generated at transition interval lengths of 0.001 times the event duration for A, 0.5 times the event duration for B, 1.0 times the event duration for C, and 5.0 times the event duration for D. Background DCA-1 value is -0.5 and completeness is 100%. Excursion magnitude (y-axis) indicates the difference between the simulated background DCA-1 value and the simulated event DCA-1 value. Resolution potential (x-axis) is the event duration divided by the time represented by the sample interval.Color shading indicates the median excursion magnitude of all samples that intersect an event. Contour lines show the parameter space where 50% (dashed line), 75% (dotted line) and 95% (solid line) of events are accurately detected.

FIGURE 9. DCA-1 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 9. DCA-1 values observed when assemblages simulated at different event values are mixed with different proportions of the background assemblage. For A, the maximum DCA-1 value observed is shown; for B, the median DCA-1 value observed; and for C, the minimum DCA-1 value observed. In all cases, background assemblages are simulated at a DCA-1 value of -0.5 and are mixed with an event assemblage with a DCA-1 value as given on the x-axis.

FIGURE 3 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 3. Multivariate comparison of empirical and simulated foraminiferal assemblages. A. Detrended correspondence analysis showing empirical (red filled) and simulated (black open) assemblages in the same ordination space. B. DCA1 scores for empirical samples paired with the DCA1 score of the corresponding simulated sample. Dotted line is the 1:1 line and is largely obscured by the points. C. All pairwise dissimilarities among empirical samples plotted against the average pairwise dissimilarity of 300 corresponding samples simulated at the same DCA1 values. Color scale depicts the density of dissimilarities with higher concentrations of dissimilarities in brighter colors. Black line is the 1:1 line.

FIGURE 7 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 7. The interaction of bioturbation, completeness and transition duration. All simulations figured include bioturbation and use a background DCA-1 value of -0.5. Simulation results for 100% sampling completeness and transition durations 0.001 times the event duration for A; 100% sampling completeness and transition durations five times the event duration for B; 25% sampling completeness and transition durations 0.001 times the event duration for C; 25% sampling completeness and transition durations five times the event duration for D. Excursion magnitude (yaxis) indicates the difference between the simulated background DCA-1 value and the simulated event DCA-1 value. Resolution potential (x-axis) is the event duration divided by the time represented by the sample interval. Color shading indicates the median excursion magnitude of all samples that intersect an event. Contour lines show the parameter space where 50% (dashed line), 75% (dotted line) and 95% (solid line) of events are accurately detected.

FIGURE 14 in Simulating our ability to accurately detect abrupt changes in assemblage-based paleoenvironmental proxies

FIGURE 14. Estimated ability of the record to detect 100 year-long events with 3 cm samples, given a known sedimentation rate at three excursion magnitudes (0.4, 1.2 and 2.8), based on 'worst-case scenario' simulations with 25% sampling, bioturbation, rapid transitions between events and the background condition (-0.5 DCA-1 value). A and B are simulated with an excursion magnitude of 0.4; C and D are simulated with an excursion magnitude of 1.2; and E and F are simulated with an excursion magnitude of 2.8. A, C and E show the percentage of events that are detectable above a background value of -0.5. An event is detected if a sample has an observed DCA-1 value exceeding the 95% quantile of the DCA-1 value of assemblages simulated at the background value. B, D and F show the median DCA-1 values recovered from samples intersecting simulated events. Values that fall above the blue dashed line are detected; that is, they exceed the 95% quantile of the DCA-1 value of assemblages simulated at the background value. Values that fall above the red dotted line are accurately detected; that is, they exceed the 2.75% quantile on DCA-1 values recovered from simulations at the event value.

opencc-by-4.0Feb 2018View details →

Dataset from "Combined Landsat and L-Band SAR Data Improves Land Cover Classification and Change Detection in Dynamic Tropical Landscapes"

These are the output land cover and land cover change raster maps from the paper, "<a href="https://doi.org/10.3390/rs10020306">Combined Landsat and L-Band SAR Data Improves Land Cover Classification and Change Detection in Dynamic Tropical Landscapes</a>," published in Remote Sensing journal. ABSTRACT. Robust quantitative estimates of land use and land cover change are necessary to develop policy solutions and interventions aimed towards sustainable land management. Here, we evaluated the combination of Landsat and L-band Synthetic Aperture Radar (SAR) data to estimate land use/cover change in the dynamic tropical landscape of Tanintharyi, southern Myanmar. We classified Landsat and L-band SAR data, specifically Japan Earth Resources Satellite (JERS-1) and Advanced Land Observing Satellite-2 Phased Array L-band Synthetic Aperture Radar-2 (ALOS-2/PALSAR-2), using Random Forests classifier to map and quantify land use/cover change transitions between 1995 and 2015 in the Tanintharyi Region. We compared the classification accuracies of single versus combined sensor data, and assessed contributions of optical and radar layers to classification accuracy. Combined Landsat and L-band SAR data produced the best overall classification accuracies (92.96% to 93.83%), outperforming individual sensor data (91.20% to 91.93% for Landsat-only; 56.01% to 71.43% for SAR-only). Radar layers, particularly SAR-derived textures, were influential predictors for land cover classification, together with optical layers. Landscape change was extensive (16,490 km2; 39% of total area), as well as total forest conversion into agricultural plantations (3,214 km2). Gross forest loss (5,133 km2) in 1995 was largely from conversion to shrubs/orchards and tree (oil palm, rubber) plantations, and gross gains in oil palm (5,471 km2) and rubber (4,025 km2) plantations by 2015 were mainly from conversion of shrubs/orchards and forests. Analysis of combined Landsat and L-band SAR data provides an improved understanding of the associated drivers of agricultural plantation expansion and the dynamics of land use/cover change in tropical forest landscapes.

opencc-by-4.0Jun 2019View details →

Planetary Surface Features Change Detection Dataset

Summary This dataset contains bi-temporal images pairs with associated labels of change or no-change describing whether there was a change in surface features between the two images acquired over the same location at two different times. This dataset contains images from four different instruments orbiting three different planets, each of which contains four representations of the bi-temporal image pair: composite grayscale, absolute difference, signed difference, and autoencoder bottleneck representations (described in detail in [1]). We also include the grayscale image tiles these representational datasets were created from. All datasets contain 100x100 images tiles that were cropped from larger images. We describe each below.  Contents hirise_rsl.zip : all subdirectories contain change and no-change examples represented as composite grayscale, absolute difference, signed difference, and autoencoder bottleneck from a before and after HiRISE image of recurring slope lineae on Mars. Subdirectory names are garni_XXXXXX_YYYYYY where Garni is the name of the crater on Mars shown in the images, XXXXXX is the HiRISE image ID of the before image, and YYYYYY is the HiRISE image ID of the after image. The "*_lcn" ending on some directories indicates that local contrast normalization was applied. The "*_gs_illum" and "*_gs_slope" directories contain composite grayscale representation with a third band that contains the difference between illumination (illum) and slope values at the same locations. Images with line endings _vflip, _hflip, _rot90, _rot180, and _rot270 were the result of vertical flips, horizontal flips, 90-deg rotations, 180-deg rotations, and 270-deg rotations of the image with the corresponding prefix. ctx_impacts.zip : all subdirectories contain change and no-change examples represented as composite grayscale, absolute difference, signed difference, and autoencoder bottleneck from a before and after CTX image of meteorite impacts on Mars. The prefix in each image name corresponds to the image pair described in the Appendix in [1]. lroc_impacts.zip : all subdirectories contain change and no-change examples represented as composite grayscale, absolute difference, signed difference, and autoencoder bottleneck from a before and after LROC image of meteorite and spacecraft landing impacts on the Moon. Filenames correspond to pair names provided in the Appendix in [1]. planet_misc.zip : all subdirectories contain change and no-change examples represented as composite grayscale, absolute difference, signed difference, and autoencoder bottleneck from a before and after PlanetScope image of miscellaneous processes on Earth. Filenames correspond to pair names provided in the Appendix in [1]. *_before_after_grayscale.zip : before and after grayscale tiles used to create image representations in above directories (indicated with _before and _after suffix in filenames). Images that contain "_0_" in the filename have the label no-change and images with "_1_" in the filename have the label change. [1] Kerner et al. (2019) Deep Learning Methods Toward Generalized Change Detection on Planetary Surfaces. In review at Journal of Selected Topics in Earth Observations and Remote Sensing. Attribution If you use this dataset in your own work, please cite this DOI: 10.5281/zenodo.2373798 as well as the paper below: Kerner, H. R., Wagstaff, K. L., Bue, B. D., Gray, P. C., Bell, J. F., & Amor, H. B. (2019). Toward generalized change detection on planetary surfaces with convolutional autoencoders and transfer learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(10), 3900-3918.

Figure 8 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 8. Evolution of I(q) vs. q upon drying for (a) Rivecourt and (b) Angeac. The inset is a close-up of the area between 0.01 and 0.07 µm−1.

Figure 5 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 5. Evolution of I(q) vs. q upon wetting for (a) Rivecourt and (b) Angeac. The inset is a close-up of the area between 0.01 and 0.07 µm−1.

Figure 3 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 3. Evolution with time of average grey levels in the wetting experiment for (a) Rivecourt and (b) Angeac. (c) Evolution with time of normalized grey levels. Blue: Rivecourt sample. Red: Angeac sample (see text for details).

Figure 2 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 2. Radiographic images of sample upon sorption experiments. The three pictures on the top are from the Angeac sample, while the four on the bottom are that of Rivecourt. The scale represents 1 cm. The marked areas correspond to the zones used for measuring average grey levels.

Figure 1 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 1. (a) Picture of neutron guide through experimental chamber. (b) Sorption experiment setup and radiographic image obtained. Wood samples were placed in an aluminum cup filled with water. Water appears dark, while aluminum is transparent to neutrons. (c) Desorption experiment setup and radiographic image obtained. Wood samples were wrapped in aluminum foils and placed in a tube, with direct air input (plastic tube, on top).

Figure 7 in Neutron imaging investigation of fossil woods: non-destructive characterization of microstructure and detection of in situ changes as occurring in museum cabinets

Figure 7. Evolution with time of average grey levels in the drying experiment for (a) Rivecourt and (b) Angeac. (c) Evolution with time of normalized grey levels. Blue: Rivecourt sample. Red: Angeac sample (see text for details). (d) Evolution of average grey levels in the drying experiments plotted as a function of the square root of time. Blue: Rivecourt sample. Red: Angeac sample.