A scalable big data approach for remotely tracking rangeland conditions

Rangelands cover around half of Earth’s land surface and provide critical ecosystem services, livelihoods for 2.7 billion people, and habitat for thousands of threatened species. Increasing pressures from unsustainable grazing practices, climate change, invasive species, and development have driven widespread degradation, undermining progress toward multiple SDGs. Effective, scalable monitoring of rangeland condition is needed to inform management and policy. Existing remote sensing platforms entail trade-offs between spatial resolution, temporal depth, and geographic coverage, limiting their transferability beyond data-rich regions. The study aims to develop a flexible, robust, and scalable approach that leverages long-term, globally available satellite archives to (1) establish a pre-1990 initial baseline of rangeland condition, (2) quantify the likelihood of condition changes (increase, stable, decline) at 30 m resolution over multi-decadal periods, and (3) separate management impacts from climate variability to evaluate the effectiveness of rangeland management interventions, demonstrated in Mongolia.

Prior efforts include the US Rangeland Analysis Platform (US RAP), which uses Landsat to provide 30 m annual maps from 1984–2017 across the USA, but relies on extensive field data for training, limiting global applicability. The Australian Rangeland and Pasture Productivity (Aus RAPP) uses MODIS for broader coverage but coarser (500 m) resolution and shorter records (since 2000). Other global approaches project rangeland change at very coarse resolutions (>25 km), while high-resolution techniques have advanced for other ecosystems (forests, croplands, wetlands) but not rangelands. Traditional approaches often use binary thresholds (e.g., p-values) and may not robustly handle high inter- and intra-annual variability in rangeland systems. Residual trend analyses (e.g., RESTREND) can separate climate effects but are unreliable when change exceeds 20%. These limitations motivate a likelihood-based, transferable framework that uses globally available data, robust model selection, and explicit handling of climate and management effects.

Study area: Mongolia, where rangelands cover ~80% of the country (>120 million ha). Management contexts include National Protected Areas (NPA), Local Protected Areas (LPA), and Community-Based Organization (CBO) areas. Data and platform: Implemented in Google Earth Engine (GEE). Primary data include Landsat surface reflectance (30 m, 1986–2020; TM/ETM+/OLI), CHIRPS precipitation (0.05°, 1986–2020), ESA CCI Land Cover (300 m, 1992–2020), soil units (SoilGrids, 250 m), terrain (Hammond landforms, 250 m), and Koppen–Geiger climate zones (1 km). Protected area boundaries and dates from TNC Mongolia; national field observations (1040 sites, 2016) from NAMEM. Vegetation index: Annual medoid/median EVI2 composites computed from Landsat after QA-based masking. EVI2 chosen over EVI to avoid blue-band artifacts in complex terrains; EVI2 is sensitive to above-ground biomass in Mongolian rangelands. Initial condition (pre-1990 baseline): Defined Land Capability Classes (LCCs) via k-means clustering over Asia using land cover, soils, terrain, and climate to group areas with similar biogeophysical capability and climate. Used Local Net Production Scaling (LNS) on mean EVI2 (1986–1990) within each LCC. Pixels at the 10th percentile set to LNS=0 (minimum performance), 90th percentile to LNS=100 (maximum performance); intermediate values linearly scaled. Pixels with LNS<=10 considered already severely degraded at baseline; for these, any later modeled stable condition was treated as persistent degradation (belief in decline set to 100%). Rangeland condition models: At each 30 m pixel, three linear models represent 20% increase, stable (0% change), or 20% decline over a window. Predicted EVI2: y = ym × (1 + t × z), where ym is window mean EVI2, t is time (years), and z is +0.20, 0, or −0.20. Threshold z is configurable by end users. Moving-window likelihoods: A 10-year moving window (sliding annually) was applied to each pixel’s EVI2 time series to generate annual trajectories (1995–2020) of model support. For each window, Akaike Information Criterion (AIC) was used to compute weights for the three candidate models. The AIC weights (summing to 1) are interpreted as beliefs (likelihoods) in increase, stable, or decline conditions. Products include windowed belief time series and overall beliefs over 1986–2020. Land cover transitions (1992–2020) were identified and areas of conversion excluded from management impact assessment. Separating climate from management: Management impacts were quantified as the difference in belief in increase between pre- and post-management periods for each protected area. To remove inter-annual climate variability (notably rainfall), the study used LCC-based reference sites across Asia: belief in increase over managed areas was compared to the average belief over all reference pixels in the same LCC over matched periods, and the reference value subtracted (i.e., rainfall-corrected management signature). Validation: Field observations (2016) classified into five degradation levels (Healthy, Slight, Moderate, Heavy, Full) using Ecological Site Descriptions criteria. Qualitative validation assessed spatial concordance between beliefs in decline and observed degradation. Quantitative validation used a confusion matrix by collapsing classes to Healthy (field: Healthy+Slight; map: Increase or Stable) vs Degradation (field: Moderate+Heavy+Full; map: Decline), reporting UA, PA, OA, and BA. Implementation assets and code are globally available; GEE code archived on Zenodo (doi:10.5281/zenodo.10806820).

• Initial condition: LCC-based benchmarking (1986–1990) showed near-maximal productivity in central Mongolia and reduced productivity in parts of the north and south. Land cover transitions (1992–2020) were limited, mostly minor grassland loss in the north.
• Trajectories of change (1986–2020): Central Mongolia, with higher population and grazing intensity, exhibited higher beliefs in decline; southern desert regions, driven more by climate variability, showed higher beliefs in increase. A decadal moving-window time series depicted a marked jump in belief in increase around 2012 in a protected area, consistent with a management intervention.
• Management impacts on LPAs: 74% of LPAs showed higher belief in increase after management when climate variability was not removed. After rainfall correction using LCC references, only ~25% retained higher belief in increase, indicating that about half of LPAs’ apparent improvements were rainfall-driven and ~25% showed no significant improvement.
• NPAs vs LPAs: 100% of NPAs showed improvement considering combined climate and management effects; after rainfall correction, 77% showed improvement attributable to management alone. NPAs outperformed LPAs with higher median belief differences and lower variance; earlier establishment dates were associated with stronger positive outcomes.
• NPA sub-strategies: Effectiveness ranking (median belief difference after rainfall correction): Strictly Protected Areas highest, followed by National Parks; Nature Reserves and Natural and Historical Monument Areas were lower. CBO areas showed improvements attributable solely to rainfall, likely due to very recent establishment (late 2018 and 2020).
• Climate context: Precipitation increased across much of Mongolia since 2011, contributing to greening signals and necessitating climate correction.
• Validation and accuracy: Spatial alignment showed higher beliefs in decline collocated with field-observed heavy/full degradation; healthy sites aligned with low decline beliefs. Confusion matrix (1986–2016 dominant change map): Degradation UA 76.7% (±4.20), PA 71.7% (±4.32); Healthy UA 81.8% (±2.96), PA 85.4% (±2.77); Overall Accuracy 79.9% (±2.56); Balanced Accuracy 78.5%.

The study demonstrates a scalable, data-driven framework that quantifies the likelihood of rangeland condition change at high spatial resolution, enabling differentiation between management impacts and climatic variability. In Mongolia, the approach reveals widespread decline in heavily grazed central regions post-1990, consistent with socioeconomic shifts (e.g., increased subsistence herding and tripling of livestock numbers) and corroborated by field observations. The climate-corrected assessment highlights that much of the apparent improvement in LPAs is rainfall-driven, while NPAs—especially those established earlier—show substantial management-attributable gains. Among NPA categories, stricter protection correlates with greater vegetation recovery. The AIC-based belief framework avoids arbitrary thresholds, accommodates varying change magnitudes, and provides interpretable likelihoods for decision support. These findings validate the utility of combining long-term Landsat data with robust statistical model selection to inform policy, prioritize interventions, and track sustainability outcomes in rangelands.

The paper introduces a globally scalable approach that integrates long-term Landsat data, LCC-based benchmarking, and AIC model selection to track rangeland condition changes and quantify management impacts at 30 m resolution over four decades. Applied to Mongolia, it identified initial conditions, mapped spatiotemporal changes, and separated climate variability from management effects, revealing significant management benefits in NPAs (especially stricter categories) and rainfall-driven improvements across many LPAs. The framework’s flexibility allows end users to adjust change thresholds, model forms, and assessment windows, and to incorporate higher-resolution sensors (e.g., Sentinel-2) and additional indicators as they become available. Future work should refine biogeophysical stratification beyond LCC, integrate field data on woody encroachment and invasive species, and extend analyses globally to support targeted, evidence-based rangeland management.

• Classification uncertainties: Lower accuracy for the Degradation class reflects challenges detecting early-stage degradation and confounding effects of human activity and climate variability. The chosen 20% change threshold is a pragmatic default and may require local adjustment.
• LCC constraints: LCC is derived from agricultural suitability proxies (soils, terrain, climate, land cover) and may not fully capture ecosystem function or vegetation community dynamics; improvements or alternative stratifications are desirable as global datasets evolve.
• Indicator specificity: Using greenness (EVI2) emphasizes herbaceous cover and may underdetect woody encroachment or invasive species dynamics; integrating fractional cover, structure metrics, or species/invasive data would broaden sensitivity.
• Temporal confounding: Time since management establishment influences outcomes; recent interventions (e.g., CBOs) may not yet manifest detectable effects.
• Land cover change handling: Although conversions were identified and excluded from management assessment, residual misclassification or unobserved changes could affect estimates.
• Data and methodological generalizability: Rainfall correction via LCC references assumes comparable climate responses within classes; deviations may introduce bias.