U.S. Geological Survey Matthew Rigge Brett Bunde Kory Postma Hua Shi 20240117 Raster Digital Data Set Earth Resources Observation and Science Center, Sioux Falls, SD U.S. Geological Survey Rigge, M., H. Shi, C. Homer, P. Danielson, and B. Granneman. 2019. Long-term trajectories of fractional component change in the Northern Great Basin, USA. Ecosphere: e02762. Rigge, M., C. Homer, L. Cleeves, D. Meyer, B. Bunde, H. Shi, G. Xian, and M. Bobo. 2020. Quantifying western U.S. rangelands as fractional components with multi-resolution remote sensing and in situ data. Remote Sensing. 12: 412. Rigge, M., Homer, C., Shi, H., Meyer, D., Bunde, B., Granneman, B., Postma, K., Danielson, P., Case, A., and Xian, G. 2021. Rangeland Fractional Components Across the Western United States from 1985 to 2018. Remote Sensing, 13: 813. https://www.mrlc.gov/data https://doi.org/10.5066/P9SJXUI1 Matthew Rigge Hua Shi Kory Postma Brett Bunde 20220805 Publication (Journal) GIScience and Remote Sensing 59:1, 1243-1265 https://doi.org/10.1080/15481603.2022.2104786 The RCMAP (Rangeland Condition Monitoring Assessment and Projection) dataset quantifies the percent cover of rangeland components across western North America using Landsat imagery from 1985-2023. The RCMAP product suite consists of ten fractional components: annual herbaceous, bare ground, herbaceous, litter, non-sagebrush shrub, perennial herbaceous, sagebrush, shrub, tree, and shrub height in addition to the temporal trends of each component. Several enhancements were made to the RCMAP process relative to prior generations. First, high-resolution training was revised using an improved neural-net classifier and modelling approach. These data serve as foundation to the RCMAP approach. The training database was further improved by incorporating additional datasets. Next, the Landsat compositing approach was improved to better capture the range of conditions from across each year and through time. These composites are based on Collection 2 Landsat data with improved geolocation accuracy and dynamic range. Finally, the Canadian portion of the sagebrush biome was included, which expanded the study area by 29,199 km2. Processing efficiency has been increased using open-source software and USGS High-Performance Computing (HPC) resources. The mapping area included eight regions which were subsequently mosaicked. These data can be used to answer critical questions regarding the influence of climate change and the suitability of management practices. Component products can be downloaded https://www.mrlc.gov/data. The goal of RCMAP is to provide a Landsat imagery-based time series of rangeland fractional components across Western North America from 1985 to 2023. These data will provide an inventory of land cover validated products with estimates of precision for the western rangelands. Climate change, shifting fire regimes, and management practices are increasingly impacting the health of the ecosystem. This dataset contributes to improved monitoring of rangeland change at broad temporal and spatial extents. Components are defined as: Bare Ground is a continuous field component including exposed soil, sand, and rocks. Annual Herbaceous is a continuous field component including grasses and forbs whose life history is complete in one growing season. This component is primarily dominated by annual invasive species including Cheatgrass (Bromus tectorum), Medusahead (Taeniatherum caput-medusae), Red Brome (Bromus rebens), or annual mustards such as Tumble Mustard (Sisymbrium altissimum) and Tansy Mustard (Descurainia pinnata), but it may contain substantial native annual herbaceous vegetation at higher elevations and in California. This component is nested within Herbaceous as a secondary component. Herbaceous is a continuous field component consisting of grasses, forbs and cacti which were photosynthetically active at any point in the year of mapping. Non-sagebrush shrub is a continuous field component encompassing all shrub species not of the sagebrush (Artemisia spp.) genus. Shrubs, in general, are discriminated by the presence of woody stems and less than 6-m in height. Perennial herbaceous is a continuous field component consisting of grasses, forbs and cacti which were photosynthetically active at any point in the year of mapping and whose lifecycle includes more than one growing season (includes biennials). Litter is a continuous field component including dead standing woody vegetation, detached plant organic matter and biological soil crusts. Sagebrush is a continuous field component encompassing almost all species of Sagebrush (Artemisia spp.) including Big Sagebrush (A. tridentata spp.), Low Sagebrush (A. arbuscula), Black Sagebrush (A. nova), Three-tip Sagebrush (A. triparta) and Silver Sagebrush (A. cana). This component is nested within Shrub as a secondary component. Excludes the low stature prairie sage (A. frigida) and white sagebrush (A. ludoviciana). Shrub is a continuous field component encompassing all shrub species discriminated by the presence of woody stems and less than 6-m in height. Tree cover is defined as vegetation with persistent woody stems greater than 6m in height. Mature stand of pinyon (Pinus spp. and juniper (Juniperus spp.) are included regardless of height. Shrub height is the average height of all shrub in centimeters. This component only occurs where the shrub cover component is greater than 0% Height is given for the portion of pixel with shrubs present. For example, in a pixel where shrub cover is 10% and the average height of those shrubs is 100cm, height will be given as 100cm, not 100cm/10% cover as 10cm. Although this Federal Geographic Data Committee-compliant metadata file is intended to document the data set in nonproprietary form, as well as in Esri format, this metadata file may include some Esri-specific terminology. 19850101 20230927 publication date Complete As needed -128.0026 -99.6758 51.5761 26.5157 ISO 19115 Topic Category biota environment geoscientificInformation imageryBaseMapsEarthCover USGS Thesaurus shrubland ecosystems terrestrial ecosystems Alexandria Digital Library Feature Type Thesaurus shrublands time series back-in-time trends grassland change shrubland change vegetation change climate change rangeland management None shrub sagebrush big sagebrush herbaceous annual herbaceous litter grass vegetation bare ground rangeland shrubland shrub height vegetation height USGS Metadata Identifier USGS:656f1f85d34ecaabeadac998 Common geographic areas United States Washington Oregon Montana North Dakota South Dakota Wyoming Idaho Nebraska Nevada Utah California Colorado Arizona Texas New Mexico Plateau Desert Northern Great Salt Lake Desert Southern Great Salt Lake Desert North Plains Plains The Rockies Canada Alberta Saskatchewan None NM TX AZ CO CA UT NV NE ID WY SD ND MT OR WA Great Basin Arizona Plateau Black Hills Blue Mountains Chihuahuan Desert Colorado Plateau Columbia Plateau Grand Canyon Middle Rockies Rocky Mountains Gunnison Sonoran Desert Southwest Tablelands Three Forks Wasatch Western US Yellowstone Northern Mountainous Mediterranean California Northern Great Plains Northern Rocky Mountains Wyoming Basin Mojave Sonoran Chihuahuan Southern Rocky Mountains Sierra Nevada mts AB SK Prairie Provinces Any downloading and use of these data signifies a user's agreement to comprehension and compliance of the USGS Standard Disclaimer. Ensure all portions of metadata are read and clearly understood before using these data to protect both user and USGS interests. There is no guarantee of warranty concerning the accuracy of the data. Users should be aware that these data were developed from models which can contain some local error. Users should not use these data for critical applications without a full awareness of their limitations. Acknowledgement of the originating agencies would be appreciated in products derived from these data. Any user who modifies these data is obligated to describe the types of modifications they perform. User specifically agrees not to misrepresent the data, nor to imply that changes made were approved or endorsed by the U.S. Geological Survey. Please refer to https://www.usgs.gov/privacy.html for the USGS disclaimer. U.S. Geological Survey Customer Services Representative mailing

47914 252nd Street

Sioux Falls SD 57198-0001 U.S. 605-594-6151 605-594-6589 custserv@usgs.gov U.S. Geological Survey (USGS) Bureau of Land Management (BLM) Environment as of Metadata Creation: Microsoft [Unknown] Version 6.2 (Build 9200); Esri ArcGIS 10.6.1 (Build 9270) Service Pack N/A (Build N/A) Jeremy D. Maestas Steven B. Campbell Jeanne C. Chambers Mike Pellant Richard F. Miller 201606 publication https://www.sciencedirect.com/science/article/pii/S0190052816000109?via%3Dihub https://doi.org/10.1016/j.rala.2016.02.002 Jon E. Keeley Sterling C. Keeley 19810401 publication https://bsapubs.onlinelibrary.wiley.com/doi/abs/10.1002/j.1537-2197.1981.tb07796.x https://doi.org/10.1002/j.1537-2197.1981.tb07796.x Emanuel A. Storey Douglas A. Stow John F. O'Leary 20160915 publication https://www.sciencedirect.com/science/article/pii/S0034425716302176 https://doi.org/10.1016/j.rse.2016.05.018 Maps are rigorously validated using field data not included as training (i.e., independent) with data from long-term monitoring sites, and by assessing model fit to training data. The independent data consists of 1) 2,014 points, each specifically designed to represent a single Landsat pixel, collected from 2013-2023, and 2) long-term monitoring data in southwest Wyoming at 126 plots observed 12 times between 2008 and 2021. The spatial-temporal correlation (n = 1,137) across all 126 Wyoming plots showed robust correlations for all components, with an R2 of 0.66 for bare ground (RMSE = 14.7%) and 0.41 for shrub cover (RMSE = 8.7%) and average R2 of 0.42 and RMSE of 9.8% across components. Next, RCMAP data were compared to independent validation sites (n = 2,014) collected from 2013-2023. Correlations between RCMAP and independent validation sites were again robust across all components, with an R2 of 0.71 for bare ground (RMSE = 14.6%) and 0.42 for shrub cover (RMSE = 10.1%) and average R2 of 0.52 and RMSE of 10.6 across components. Bare Ground - R2 0.71, RMSE 14.6, Herbaceous - R2 0.65, RMSE 11.5, Litter - R2 0.36, RMSE 9.6, Shrub - R2 0.42, RMSE 10.1, Sagebrush - R2 0.43, RMSE 7.1, Annual Herbaceous - R2 0.52, RMSE 10.1. Next, RCMAP data were compared to a 5% withholding of BLM AIM and LMF data not used for training (n = 6,539) collected between 2011 and 2022. Correlations between RCMAP and AIM/LMF data were again robust across all components, with an R2 of 0.61 for bare ground, R2 of 0.67 for tree, and 0.38 for shrub cover and average R2 of 0.41 and RMSE of 11.8 across components. Bare Ground - R2 0.61, RMSE 13.1, Herbaceous - R2 0.54, RMSE 17.0, Litter - R2 0.04, RMSE 12.5, Shrub - R2 0.38, RMSE 10.6, Sagebrush - R2 0.42, RMSE 8.0, Annual Herbaceous - R2 0.39, RMSE 13.3, Tree - R2 0.67, RMSE 8.2, Shrub Height - R2 0.13, RMSE 69.4. Finally, the cross-validation of predictions were evaluated against training data at the Landsat scale. Cross-validation correlations included an R2 of 0.92 for bare ground (RMSE = 8.1%) and 0.69 for shrub cover (RMSE = 9.1%) and average R2 of 0.81 and RMSE of 6.9 across components. Bare Ground - R2 0.92, RMSE 8.1, Herbaceous - R2 0.87, RMSE 8.5, Litter - R2 0.75, RMSE 4.2, Shrub - R2 0.69, RMSE 9.1, Sagebrush - R2 0.74, RMSE 3.9, Annual Herbaceous - R2 0.83, RMSE 5.8, Tree - R2 0.87, RMSE 9.1, and shrub height – R2 0.72, RMSE 16.0. Changes observed in both the field and RCMAP data were typically gradual, within-state, changes which are most difficult to resolve, which were often successfully captured. It is important to consider that all accuracy assessments described above are designed to evaluate single-pixel level correspondence. Due to fine-scale landscape heterogeneity this is the most rigorous approach, and most applications looking at broader spatial scales would tend to lower error relative to this analysis. The methods employed to map fractional vegetation components in rangeland ecosystems in the Western U.S. include modelling rangelands as a series of independent continuous field components from 1985 – 2023 using field observations, neural network classifiers, and multiple Landsat composites. This fractional estimation of ten rangeland habitat variables ranging from 1985 - 2023 in the Western North America is the version dated January 2024. Data set is considered complete for the information presented, as described in the abstract. Users are advised to read the rest of the metadata record carefully for additional details. A formal accuracy assessment of the horizontal positional information in the data set has not been conducted. A formal accuracy assessment of the vertical positional information in the data set has either not been conducted or is not applicable. The RCMAP (Rangeland Condition Monitoring Assessment and Projection) dataset quantifies the percent cover of rangeland components across western North America using Landsat imagery from 1985-2023. The RCMAP product suite consists of ten fractional components: annual herbaceous, bare ground, herbaceous, litter, non-sagebrush shrub, perennial herbaceous, sagebrush, shrub, tree, and shrub height and the temporal trends of each. The five primary components, bare ground, shrub, litter, herbaceous, and tree are designed to sum to 100% in each pixel. The secondary components of annual herbaceous and perennial herbaceous and non-sagebrush shrub and sagebrush are subsets of the primary components herbaceous and shrub, respectively. Processing occurred in eight regions which were subsequently mosaicked across the study area. The eight regions: Mediterranean California, Great Basin and Columbia Plateau, Northern Rocky Mountains, Northern Great Plains and Wyoming Basin, Warm Deserts (Mojave, Sonoran, and Chihuahuan), Southern Rocky Mountains and Sierra Nevada, Colorado Plateau and Southwest Tablelands, and Pacific Northwest. 2023 High-resolution training data – Previous iterations of RCMAP have relied on Cubist classifications trained on each image footprint with a sample size of ~95 per image. The classifier and training paradigm was revised based on optimization analysis. Neural Net and Convolutional Neural Net classifications were tested as a replacement for Cubist, finding the best results with Neural Nets; yielding an accuracy on average 15% better than legacy cubist runs. Keras tuner was also tested, and it was found that it did not improve results beyond our preset parameters. Finally, a 3 by 3-pixel variance was added for each high-resolution image band to the predictor stack significantly improved accuracy. Specifically, a higher variance was found associated with higher shrub and tree cover, a pattern successfully leveraged by the Neural Network model. Next, the impact of pooling multiple image sites was investigated from within similar ecological conditions (same level 3 ecoregion) into regional models compared to individual models for each image as done in prior versions. The imagery was split into 15 Worldview (WV) groups (average of 20.1 images per group), 4 Pleiades groups (10.5 images per group), and 1 QuickBird group (5 images). These image groups had an average of 2,106 training polygons (22,451 pixels), maximum of 4,907 (54,073 pixels), and minimum of 375 (1,440 pixels), compared to an average of 95 plots per image. Pooling data into regional models offers the advantages of increasing the size and range of training data and spectral conditions (Okin et al. 2001, Xu et al. 2016) and improving regional consistency. While pooling data into ecoregion-based models can regress results toward the mean (Zhou et al. 2020, Kearney et al. 2022), it was found on average 15% improvement in high-resolution model accuracy in the pooled models as compared to independent runs, a pattern consistent with Zhou and Kearney. Processing regionally also allows the addition of 18 WV images where no field data were collected do to access or other logistical challenges. These collects with a viable image, but no field data, had component cover predictions generated from the data available from sites within the same ecological grouping. Moreover, the training footprint was expanded by no longer excluding NLCD Hay/pasture areas. Together, these changes resulted in a 45% increase in the high-resolution training footprint to 58,049 km2. To improve consistency among regions some image sites and corresponding field data near ecotones were used to train multiple regions, and predictions for these sites were averaged across model runs. Independent variables for the high-resolution models included imagery, 3-by-3 focal coefficient of variation, indices (NDVI, red/green, NIR/yellow), topographic elevation downscaled to 2-m, latitude, and longitude. Our final model was a DNN with a neuron width of 256, 4-layers deep, learning rate of 10^-4, dropout between layers of 0.2, batch normalization, and clamping of 0-100% (0-500 cm for shrub height). A weighted mean squared logarithmic error (WMSLE) loss function was used, where zeros have half the weight of nonzero values to reduce the abundance of zeros in predictions. The point of the WMSLE approach is to penalize higher relative errors, for example a prediction of 5% cover with a training value of 0% is given a higher penalty than a prediction of 50% cover with a training value of 55%. Whereas commonly used loss functions such as MAE would consider the above example to be equivalent. A model was developed with all cover components and used the subsequent classifications as independent variables in the shrub height model. It was found that adding cover components as variables significantly improved shrub height classification accuracy as they, especially shrub cover, are strongly related to height. To test high-resolution model performance, a 5% sample was utilized. The 5% sample was selected randomly from a pool of plots greater than 90 m away from other sites (representing 41% of the total pool) to minimize spatial autocorrelation bias (e.g., Macander et al. 2022). Cross validation accuracy (R2) was 0.73 for annual herbaceous, 0.83 for herbaceous, 0.66 for litter, 0.71 for sagebrush, 0.69 for shrub, 0.91 for bare ground, 0.89 for tree, 0.61 for shrub height, with and overall average of 0.75. Independent accuracy results were on average 26% weaker than cross validation. Spatial patterns in the high-resolution predictions closely matched expected patterns observed in the field. Following initial high-resolution classification, a series of post-processing steps was utilized to improve accuracy. 1) Since field plots in high-resolution sites had minimal tree cover, tree cover training values were extracted from a Unet-based tree cover classification with significant hand editing for each site. 100 tree cover training points (all at 100% cover) were added for each collect in the processing region. 2) Field plots contained minimal samples on hay, pasture, or otherwise irrigated land covers, resulting in overestimation of tree, shrub, and sagebrush cover over these areas. To remedy this issue, areas in which no woody cover is likely to occur (NLCD 2021 Hay/pasture class and non-woody Cropland Data Layer [CDL] crop type) were identified. In these areas, in the initial prediction the values for shrub, sage, and tree cover, were set to zero and proportionally added their cover to herbaceous, annual herbaceous, litter, and bare ground. Next, 100 training points for each collect in the processing region were added from the Hay/pasture extent, with the proportioned training data as input. 3) To fully leverage our Unet-based classification of trees, a convergence of evidence approach was used to label shrub and tree. Specifically, in pixels identified as tree in the Unet product, the shrub cover prediction was set to zero and added the initial prediction to tree cover. Conversely, in pixels identified as non-tree, the tree cover prediction was set to zero and added the initial prediction to shrub cover. Finally, in pixels sharing any boundary with a tree classified pixel, the initial shrub and tree predictions were kept. The 2 m predictions of component cover were resampled to 30 m using bilinear interpolation. Next, at the 30 m scale the sum of primary components was ensured as 100% and the rectification was intact (shrub greater than sage, herbaceous greater than annual herbaceous, and shrub height was greater than 0 only if shrub cover was greater than 0. Since the collect data are based on ocular estimates, which have known bias relative to objective measurements (e.g., line point intercept) (Murphy and Lodge 2002, Abbott 2009), a correction function was applied to our data: visual cover = 0.925 X objective cover ^2). Specifically, the bias addresses a deflation in values in the mid-ranges (~20-80%) of visual data relative to objective data. Finally, to ensure proper training of burned areas, the data were conformed to the limits of shrub, sage, and tree cover based on time-since fire (Rigge et al. 2019), and in the current implementation, the data were additionally stratified by ecosystem resistance and resilience classes (Maestas et al. 2016). 2023 Image Processing – Collection 2 Landsat Analysis Ready Data (ARD) data provide improved spatial and radiometric qualities and more image availability, which enhance training accuracy and time-series consistency relative to collection 1 (Dwyer et al. 2018). Landsat 5, 7, 8, and 9 data including 6 surface reflectance bands from the Collection 2 ARD and corresponding Quality Assessment (QA) bands were used in the current generation of RCMAP. Moreover, the compositing approach was revised to better capture the peak greenness, the range of conditions over a year, and spatial consistency. This new approach results in improved dynamic range of greenness/brightness among composites relative to our prior approach relying on median values from regionally defined compositing windows. Moreover, this new percentile approach avoids the need to set seasonal windows (e.g., leaf-on in Northern Plains April 1-June 30), rather it dynamically selects the percentile values from across a given year, resulting in greater consistency across regions. Finally, a custom snow/cloud/shadow masking procedure was developed to augment pixel QA masking. First, a reference image was developed using the regional image composites utilized in the 2022 release (Rigge et al. 2022). Looking through all “leaf-on” images from 1985-2021, the mean minus 2 standard deviations was calculated for each pixel. Next, for each new candidate collection 2 pixel, the Normalized Burn Ratio (NBR, Lopez et al. 1991), snow index (SWIR 2 x 100) / Red, and the Cloud Shadow Index (CSI) (Zhai et al. 2018) were calculated. Pixels were flagged with a mask value if 1) candidate collection 2 values for the NIR or SWIR 1 bands were lower than the reference image for that region, 2) CSI values indicate high confidence for clouds or shadows, or 3) snow index values (less than 81) indicate the presence of snow. To remove masking commission in dark lava, bright playas, and recently burned areas, the preliminary masking in pixels with an NBR less than 0.04 were removed. Finally, cloud/snow/shadow pixels were buffered by 1 pixel. QA (pixels not marked clear or water but may be shadow or snow) and additional custom snow/shadow/cloud masking were applied to all ARD imagery. Additionally, for each pixel, the Z-score of each band relative to the ARD observation mean and standard deviation were calculated, ignoring the areas masked as described above. Pixels where the range between the maximum and minimum Z-score among bands exceeded 6 were also masked. This additional Z-score masking procedure removed image artifacts, especially near scene edges, and snow cover. From the remaining clear pixels, the 10th, 50th, and 90th percentile values were calculated on a band-by-band basis for all images available within each evaluated year (1985-2023). Calculation of percentile values varied based on the number of clear observations available. 1) If pixel clear observation count was 0, the median of observations that passed the QA masking, but not custom masking was calculated. 2) If pixel clear observation count was 1 or 2, pixels greater than 2 standard deviations above or below the mean were removed, then calculated the percentile values from the remaining observations. 3) If pixel clear observation count was 3-11, pixels in the lowest and highest observations that year were removed, then calculated the percentile values from remaining observations. 4) If pixel clear observation count was greater or equal to 12, pixels in the 2 lowest and 2 highest observations that year were removed and calculated the percentile values from remaining observations. This staggered approach based on clear observation count produced the highest quality composites by removing noise (snow, shadows, cloud cover, image artifacts) often found in the highest and lowest observations within a year. The paradigm allows one to be more selective on pixel quality as clear count increases. For 2023, images were obtained through September 27. For 2012, with limited Landsat imagery, was filled with the average of 2011 and 2013, except for areas identified as burned in 2012 by MTBS, was filled with 2013 only. 2023 Ancillary Data -- In addition to Landsat composites and CCDC-synthetic imagery, several ancillary data layers were included as independent variables in the regression modelling of fractional components. 1) Topographic data; slope, aspect, elevation, and position index. In the current generation a new version of aspect data was used, processed into Cartesian (x, y) coordinates. 2) For each Landsat composite, seven spectral indices were calculated, the Soil Adjusted Vegetation Index (SAVI), Normalized Difference Built-up Index (NDBI), Normalized Difference Water Index (NDWI), and tasseled cap wetness, brightness, and greenness. 3) Geographic location information (i.e., latitude and longitude/ x and y coordinates). 4) Water bodies occurring at any point during the time-series were detected using the NDWI. Areas identified as agriculture or water at any point in the time-series were excluded from the training data pool and included in the land cover mask. 5) Monitoring Trends in Burn Severity (MTBS) and GeoMac burn data were used to produce a suite of burned area products. 2023 Land Cover Masking -- A land cover mask based on pixels identified as non-rangeland in the circa 2016 base year (Rigge et al. 2020) was applied during the process and to the products. The non-rangeland mask includes urban areas, major roads, snow, and ice identified in NLCD 2021 land cover classes in addition to cultivated crops and water. Cultivated croplands were masked using a combination of the 2013 Cropland Data Layer from the National Agricultural Statistics Service and NLCD 2021 classes. Open water areas were identified using Normalized Difference Water Index (NDWI) thresholds. NLCD hay/pasture and forest areas masked in prior generations of RCMAP data are not masked in the current generation of products. 2023 Change detection – Each normalized target and base composite pair were compared using a Change Vector (CV) approach, finding differences in spectral values by band. Spectral differences were summed across bands to produce a change vector magnitude value for each composite pair. If a pixel had a change vector magnitude beyond a standard deviation threshold in a given target year composite for a given NLCD 2021 land cover class, it was flagged as changed (Rigge et al. 2019). Though 10th, 50th, and 90th percentile Landsat composites were generated, the 10th and 90th were used for change detection. Two out of two composite agreement in change was needed for a pixel to be flagged as changed between a target and base year. The requirement for two composite agreement was critical to remove cloud, shadow, and haze influence missed by the QA masking. Next, an approach was used which considers the range of CV values through time at each pixel. The CV magnitude range represents the extent of variability through time at each pixel, those pixels with stable conditions have a low value, while those with disturbance/change have a high value. The yearly CV magnitude was compared to the range, allowing one to define the degree of confidence in change. This metric was defined as Change Fraction (CF). The CF considers three weighted attributes in change detection. First, the composite CV magnitude score for a given composite in relation to the CV magnitude range per pixel. This gives a high value to pixels with a high amount of spectral change relative to the total range of variation in the time-series. Second, the CV magnitude range per pixel is considered, relative to the maximum in each region. This gives a high value to pixels with a high total variation in the time-series relative to those pixels with lower variation through time. Third, the CV magnitude per pixel relative to the maximum CV magnitude in each region, each year. The result is a high value to pixels with a high amount of spectral change relative to the maximum change in that composite. The total score for each composite is calculated and the maximum of the two composites in a year is given as the final CF. Values of CF ranged from 0 (high confidence unchanged) to 100 (high confidence changed). In the current version, the weighting of the three attributes was revised as described above by giving more weight to the first attribute: the CV magnitude score for a given composite in relation to the CV magnitude range per pixel. Pixels with a high amount of difference in CF score between composites in a year are often indicative of haze/atmospheric contamination (in addition to real change). If difference between composites was high, then the minimum CF value was taken, otherwise the maximum value was used. 2023 Training – From the pool of ~64 million 30 m pixels from high-resolution sites, 10% available within each of our 8 processing regions were selected (Rigge 2023). Training data from the BLM AIM, LANDFIRE Public Reference Database, and Landscape Data Commons datasets were obtained. All data, aside for a random 10% withholding from the AIM database were used as common independent validation data across RCMAP, RAP, and LandCART. Finally, 1,600-line point intercept observations of component cover over southeast Alberta from 1998-2003 (Aldridge 2005) were obtained. All yearly sampled data files per component were appended into a combined spatio-temporal database. Like the points added to hay pastureland cover at the high-resolution scale, 10,000 training points were added per region in 2001 and 2019 (to correspond with NLCD epochs) to the Landsat scale training pool. As with the high-resolution scale, the initial prediction was used with zeroed out shrub, sage, and tree cover, and proportionally added their cover to herbaceous, annual herbaceous, litter, and bare ground. Finally, while high-resolution collect data included some incidental recent burns, they do not represent a systematic sample of burned history. Burned areas, especially those recently burned, tend to contain unique spectral values that do not reflect conditions in burned areas. For example, a recently burned site with 0% shrub, 30% herbaceous, 25% litter, and 45% bare ground has on average distinct spectral conditions relative to the body of such analogues in unburned areas. To improve representation of burned areas in our training pool, 30,000 points were added per year per region to pixels where moderate or severe fires defined by MTBS occurred in the current or preceding 3 years. To these points, a value of 0% cover was assigned for shrub, sage, and tree. The burned area data were appended to the main training pool, which ensures predictions greater than 0% cover of shrub and sagebrush in burned areas, especially in unburned islands (Rigge et al. 2021). Landsat scale neural net model architecture was like that described above for the high-resolution scale, though the list of independent variables differed. Notably, synthetic Landsat data are no longer included as input, as an improved compositing method was found and associated increased seasonal dynamics negated their need. Variables used; topographic data; slope, Cartesian aspect, elevation, and position index, elevation, latitude, longitude (i.e., m of northing and easting), from each Landsat composite; SAVI, NDBI, NDWI, and tasseled cap wetness, brightness, and greenness. Models were developed for each region independently. No training data with values over 0% sagebrush were available in the California region, though limited amounts of sagebrush do occur. To remedy this issue, the sagebrush model from the Great Basin region was applied to the California data yielding values up to 18% sagebrush cover. Test accuracies revealed reliable results from both models, but the model trained across the study area tended to produce more generalized (i.e., flattened) results. Two sets of predictions were generated for each region, one for recently burned areas (burned area model), that included the 30,000 points per year per region with values of 0% shrub, sage, and tree as described above, and another without these additional data (unburned area model). The burned area model results were inserted in pixels that burned in the target year and the preceding 1 year in California chapparal, 5 years in areas of high ecosystem resistance and resilience (RandR), 5 years in moderate RandR, and 10 years in low RandR. Unburned area predictions were kept elsewhere. For shrub height, like the high-resolution scale, the initial cover predictions were used as independent variables to improve performance. 2023 Component Prediction -- Using the spatio-temporal training pool with the adjusted base values a neural network model was developed and used to predict all components, using the component values as the dependent variable, and the full set of independent variables. Neural network models were developed for the spatio-temporal training dataset but applied the model to data from each target year to produce component cover predictions. A second set of predictions was run for shrub, sagebrush, and tree using training data that included recently burned areas. This burned area prediction for shrub, sagebrush, and tree was inserted in pixels that burned in recently burned areas, while the standard prediction was kept elsewhere. 2023 Change Labeling -- CF data was used to determine if pixels should be labeled with the target year estimate or if the base value should be maintained. Specifically, the criteria for change were CV = 1 and CF greater than 29 or CV = 0 and CF greater than 60. Component cover change was limited in four situations. First, if component change between a given target and base year was greater than defined thresholds described in (Rigge et al. 2019), it was lowered to the limit. Next, the primary fractional components and tree canopy cover were summed, and if necessary, the component cover values were adjusted, so the summation was 100% while maintaining the proportions of each component. Third, areas were identified in which most pixels in a 5 by 5 moving window had +/- 1% component cover change between the base and target year. Similarly, areas were identified in which most pixels had no change, and less than 20% had +/- 1% component cover change. Since the change in both above scenarios was likely to be noise, the base year fractional component cover was inserted in these changed pixels, while the target year estimate was maintained in the remaining area. Fourth, the total number of fractional components with change in each target year. If only 1 out of 9 fractional components in a pixel had change, the base year fractional component cover was maintained for all fractional components. The logic was that if one fractional component changes, then other component(s) should change in response. Finally, the fractional component relationships were insured to be intact (shrub greater than sagebrush, herbaceous greater than annual herbaceous, shrub height = 0 if shrub cover = 0). If not, the relationships were forced to be intact. 2023 Post Processing -- Post-processing has been improved with updated fire recovery equations stratified by ecosystem resistance and resilience (R and R) classes (Campbell and Maestas 2016) to stratify recovery rates. Ecosystem R and R maps are only available for the sagebrush biome. Classes with 1985-2020 average water year precipitation to identify precipitation thresholds corresponding to R and R classes were intersected. Outside of the sagebrush biome, precipitation was used to produce R and R equivalent (low, medium, high). Due to the fast recovery following fire in California chapparal (e.g., Keeley and Keeley 1981, Storey et al. 2016), EPA level 3 ecoregions were used to define a 4th R and R zone. Recovery rates are based on Arkle et al. (in press) who evaluated the recovery of plant functional groups in 1278 post-fire rehab plots by time since disturbance stratified by ecosystem resistance and resilience. This analysis was expanded by evaluating the postfire-recovery in all AIM and LMF data across the West to establish second order polynomial equations defining maximum sage, shrub, and tree cover by time-since fire by R and R class. Recovery limits in California follow Keeley and Keeley 1981 and Storey et al. 2016. Fire correction percent cover limits (y) for sagebrush by each R and R zone were; zone 1) (high R and R) y = 0.0332x2 – 0.1866x + 3.9794, zone 2) (medium R and R) y = 0.0234x2 – 0.0988x + 1.9323, zone 3) (low R and R) y = 0.00074x2 – 0.1535x + 0.676, zone 4) (California chapparal) y = 0.0332x2 – 0.1866x + 3.9794, where x is the number of years since most recent fire. Fire correction percent cover limits (y) for shrub and tree by each R and R zone were; zone 1) (high R and R) y = 0.0019x2 + 1.611x + 3.564, zone 2) (medium R and R) y = 0.0141x2 + 0.5558x + 2.540, zone 3) (low R and R) y = -0.0011x2 + 0.5891x + 1.1267, zone 4) (California chapparal) y = 0.6.7857x2 – 7.6429x + 20.571, where x is the number of years since most recent fire. Most time-series noise detection procedures are not appropriate in rangelands due to their interpretation of rapid changes related to interannual weather variation noise (Browning et al. 2018, Zhou et al. 2020). A custom noise detection model was developed. For each pixel, a third order polynomial model was fit for each component cover time-series. Observations with a z-score more than 2 standard deviations from the mean are removed, and a new third order polynomial model (i.e., cleaned fit) is fit to observations within this threshold. Finally, looking again at all observations, those observations with a z-score more than 2 standard deviations from the mean of the cleaned fit are replaced with the mean of the prior and subsequent year component cover values. These noise removal procedures were implemented in pixels that did not burn during the time-series, and in pre-burn years in pixels that did burn during the time-series. 2023 Finalizing Steps – The data were mosaiced from the eight regions, which were designed with a 1.5 km overlap. The weighting function was calculated based on distance to the edge of the overlap to feather regional predictions together. The secondary components were reconciled to the primary components for sagebrush to shrub, annual herbaceous to herbaceous, and shrub height to shrub cover. Next, perennial herbaceous and non-sagebrush shrub components were calculated by subtracting annual herbaceous from herbaceous cover and sagebrush from shrub cover, respectively. A land cover mask based on pixels identified as non-rangeland in the circa 2016 base year (Rigge et al. 2020) was applied during the process and to the products. The non-rangeland mask includes cultivated crops, water, urban areas, major roads, snow, and ice identified in NLCD 2021 and 2013 Cropland Data Layer from the National Agricultural Statistics Service. NLCD hay/pasture and forest areas masked in prior generations of RCMAP data are not masked in the current generation of products. Land cover masking and areas outside the mapping region were assigned as a no-data value (i.e., null). Finally, extent masking was applied to sagebrush and annual herbaceous components following Rigge et al. (2020). Sagebrush and annual herbaceous extent masking was typically based on thresholds combining elevation, aspect, and latitude. Specifically, these maps restricted the distribution of annual herbaceous at high elevation greater than ~2300 m and sagebrush at greater than ~2700 m. Shrub height values in locations with greater than 0, but less than ~10% shrub cover, tend to be more unreliable, and in some cases will be predicted as 0 cm height. While we corrected shrub height predictions in areas with no predicted shrub cover, we do not adjust shrub height/shrub cover in the scenario described above. 2023 Trends -- The temporal patterns were assessed in each RCMAP component with two approaches, 1) linear trends and 2) a breaks and stable states method with an 8-year temporal moving window based on structural change at the pixel level. Linear trend products include slope and p-value calculated from least squares linear regression. The slope represents the average percent cover change per year over the times-series and the p-value reflects the confidence of change in each pixel. The structural change method partitions the time-series into segments of similar slope values, with statistically significant breakpoints indicating perturbations to the prior trajectory. The break point trends analysis suite relies on structural break methods, resulting in the identification of the number and timing of breaks in the time-series, and the significance of each segment. The following statistics were produced: 1) for each component, each year, the presence/absence of breaks, 2) the slope, p-value, and standard error of the segment occurring in each year, 3) the overall model R2 (quality of model fit to the temporal profile), and 4) an index, Total Change Intensity. This index reflects the total amount of change occurring across components in that pixel. The linear and structural change methods generally agreed on patterns of change, but the latter found breaks more often, with at least one break point in most pixels. The structural change model provides more robust statistics on the significant minority of pixels with non-monotonic trends, while detrending some interannual signal potentially superfluous from a long-term perspective. These data will be part of a separate data release. 2024 Raster Grid Cell 83246 66841 1 Albers Conical Equal Area 29.5 45.5 -96.0 23.0 0.0 0.0 row and column 29.999999999836753 29.999999999810505 meters North_American_Datum_1983 GRS 1980 6378137.0 298.2572221010042 Attribute Table Table containing attribute information associated with the data set. Producer defined Value Column for value indicating per-pixel percent for bare ground, shrub, herbaceous, litter, sagebrush, annual herbaceous, non-sagebrush shrub and perennial herbaceous components range from 0 to 100. Producer defined 0 100 Percent Value (Shrub Height) In shrub height, column for value indicating per-pixel average height of all shrub in centimeters. Shrub height values greater than 0 only occur where the shrub cover component is greater than 0% Height is given for the portion of pixel with shrubs present. The shrub height component ranges from 0 to 500 cm. Producer Defined 0 500 Centimeters Count All raster attribute tables include a column for count describing a nominal integer value that designates the number of pixels that have each value in the file; histogram column in ERDAS Imagine raster attributes table. Producer defined 0 125551462 Integer Red Red color code for RGB. The value is arbitrarily assigned by the display software package, unless defined by user. Producer defined 0 100 Percent Green Green color code for RGB. The value is arbitrarily assigned by the display software package, unless defined by user. Producer defined 0 100 Percent Blue Blue color code for RGB. The value is arbitrarily assigned by the display software package, unless defined by user. Producer defined 0 100 Percent Opacity A measure of how opaque, or solid, a color is displayed in a layer. Producer defined 0 100 Percent The entity and attribute information provided here describes the tabular data associated with the data set. Please review the detailed descriptions that are provided (the individual attribute descriptions) for information on the values that appear as fields/table entries of the data set. Cell size is effectively 30-meters, but depending on software used, the data properties may show slight deviations (e.g., cell size of 29.99999999....). In this version, we adopt EPSG code 5070 as the projection, with a NAD 1983 Datum, GRS 1980 Ellipsoid, and Albers projection. Previous RCMAP iterations used the WGS 84 Datum which lacks an EPSG code in an Albers projection. The entity and attribute information were generated by the individual and/or agency identified as the originator of the data set. Please review the rest of the metadata record for additional details and information. U.S. Geological Survey GS ScienceBase mailing address

Denver Federal Center, Building 810, Mail Stop 302

Denver CO 80225 United States 1-888-275-8747 sciencebase@usgs.gov Unless otherwise stated, all data, metadata and related materials are considered to satisfy the quality standards relative to the purpose for which the data were collected. Although these data and associated metadata have been reviewed for accuracy and completeness and approved for release by the U.S. Geological Survey (USGS), no warranty expressed or implied is made regarding the display or utility of the data for other purposes, nor on all computer systems, nor shall the act of distribution constitute any such warranty. Raster Digital Data Set https://doi.org/10.5066/P9SJXUI1 None 20240117 U.S. Geological Survey Customer Services Representative mailing and physical

47914 252nd Street

Sioux Falls SD 57198-0001 U.S. 605-594-6151 605-594-6589 custserv@usgs.gov FGDC Content Standard for Digital Geospatial Metadata FGDC-STD-001-1998