Mapping land use and land cover (LULC) using remote sensing is fundamental to environmental monitoring, spatial planning and characterising drivers of change in landscapes. We develop a new, general and versatile approach for mapping LULC in landscapes with relatively gradual transition between LULC categories such as African savannas. The approach integrates a well-tested hierarchical classification system with the computationally efficient random forest (RF) classifier and produces detailed, accurate and consistent classification of structural vegetation heterogeneity and density and anthropogenic land use. We use Landsat 8 OLI imagery to illustrate this approach for the Extended Greater Masai Mara Ecosystem (EGMME) in southwestern Kenya. We stratified the landscape into eight relatively homogeneous zones, systematically inspected the imagery and randomly allocated 1,697 training sites, 556 of which were ground-truthed, proportionately to the area of each zone. We directly assessed the accuracy of the visually classified image. Accuracy was high and averaged 88.1% (80.5%–91.7%) across all the zones and 89.1% (50%–100%) across all the classes. We applied the RF classifier to randomly selected samples from the original training dataset, separately for each zone and the EGMME. We evaluated the overall and class-specific accuracy and computational efficiency using the Out-of-Bag (OOB) error. Overall accuracy (79.3%–97.4%) varied across zones but was higher whereas the class-specific accuracy (25.4%–98.1%) was lower than that for the EGMME (80.2%). The hierarchical classifier identified 35 LULC classes which we aggregated into 18 intermediate mosaics and further into five more general categories. The open grassed shrubland (21.8%), sparse shrubbed grassland (10.4%) and small-scale cultivation (13.3%) dominated at the detailed level, grassed shrubland (31.9%) and shrubbed grassland (28.9%) at the intermediate level, and grassland (35.7%), shrubland (35.3%) and woodland (12.5%) at the general level. Our granular LULC map for the EGMME is sufficiently accurate for important practical purposes such as land use spatial planning, habitat suitability assessment and temporal change detection. The extensive ground-truthing data, sample site photos and classified maps can contribute to wider validation efforts at regional to global scales.

To evaluate land use and land cover (LULC) maps an independent and representative test dataset is required. Here, a test dataset was generated via stratified random sampling approach across all areas in Fiji not used to generate training data (i.e. all Tikinas which did not contain a training data point were valid for sampling to generate the test dataset). Following equation 13 in Olofsson et al. (2014), the sample size of the test dataset was 834. This was based on a desired standard error of the overall accuracy score of 0.01 and a user's accuracy of 0.75 for all classes. The strata for sampling test samples were the eight LULC classes: water, mangrove, bare soil, urban, agriculture, grassland, shrubland, and trees.

There are different strategies for allocating samples to strata for evaluating LULC maps, as discussed by Olofsson et al. (2014). Equal allocation of samples to strata ensures coverage of rarely occurring classes and minimise the standard error of estimators of user's accuracy. However, equal allocation does not optimise the standard error of the estimator of overall accuracy. Proportional allocation of samples to strata, based on the proportion of the strata in the overall dataset, can result in rarely occurring classes being underrepresented in the test dataset. Optimal allocation of samples to strata is challenging to implement when there are multiple evaluation objectives. Olofsson et al. (2014) recommend a "simple" allocation procedure where 50 to 100 samples are allocated to rare classes and proportional allocation is used to allocate samples to the remaining majority classes. The number of samples to allocate to rare classes can be determined by iterating over different allocations and computing estimated standard errors for performance metrics. Here, the 2021 all-Fiji LULC map, minus the Tikinas used for generating training samples, was used to estimate the proportional areal coverage of each LULC class. The LULC map from 2021 was used to permit comparison with other LULC products with a 2021 layer, notably the ESA WorldCover 10m v200 2021 product.

The 2021 LULC map was dominated by the tree class (74\% of the area classified) and the remaining classes had less than 10\% coverage each. Therefore, a "simple" allocation of 100 samples to the seven minority classes and an allocation of 133 samples to the tree class was used. This ensured all the minority classes had sufficient coverage in the test set while balancing the requirement to minimise standard errors for the estimate of overall accuracy. The allocated number of test dataset points were randomly sampled within each strata and were manually labelled using 2021 annual median RGB composites from Sentinel-2 and Planet NICFI and high-resolution Google Satellite Basemaps.

Data format

The Fiji LULC test data is available in GeoJSON format in the file fiji-lulc-test-data.geojson. Each point feature has two attributes: ref_class (the LULC class manually labelled and quality checked) and strata (the strata the sampled point belongs to derived from the 2021 all-Fiji LULC map). The following integers correspond to the ref_class and strata labels:

1. water
2. mangrove
3. bare earth / rock
4. urban / impervious
5. agriculture
6. grassland
7. shrubland
8. tree

Use

When evaluating LULC maps using test data derived from a stratified sample, the nature of the stratified sampling needs to be accounted for when estimating performance metrics such as overall accuracy, user's accuracy, and producer's accuracy. This is particulary so if the strata do not match the map classes (i.e. when comparing different LULC products). Stehman (2014) provide formulas for estimating performance metrics and their standard errors when using test data with a stratified sampling structure.

To support LULC accuracy assessment a Python package has been developed which provides implementations of Stehman's (2014) formulas. The package can be installed via:

pip install lulc-validation

with documentation and examples here.

In order to compute performance metrics accounting for the stratified nature of the sample the total number of points / pixels available to be sampled in each strata must be known. For this dataset that is:

1. 1779768,
2. 3549325,
3. 541204,
4. 687659,
5. 14279258,
6. 15115599,
7. 4972515,
8. 116131948

Acknowledgements

This dataset was generated with support from a Climate Change AI Innovation Grant.

Precise land use land cover (LULC) data are essential for understanding the landscape structure and spatial pattern of land use/cover in the Yellow River Basin (YRB) to regulate scientific and rational territorial spatial planning and support sustainable development. However, differences in the multiple sets of LULC products in portraying the land composition of the YRB limit our understanding of the land cover composition in this region. To address this issue, this study chose five sets of open and high spatiotemporal LULC data in 2020, namely, CLCD, LSV10, ESRI10, CLC_FCS30, and Globeland30, to evaluate the accuracy and consistency of classification in the YRB. Our results show that: (1) The LULC composition of the YRB in 2020 was mapped consistently by the five datasets. Grasslands, croplands, and woodlands constitute the major LULC types, accounting for 96% of the total area of the study area. (2) The correlation coefficients of the LULC types of any two of the five datasets ranged from 0.926 to 0.998, showing high land compositional consistency. However, among the five datasets, there were considerable differences in the areas of a single LULC type. (3) The classification consistencies of croplands, woodlands, grasslands, and water bodies were higher than 60% in any two datasets. The spatial consistencies of grasslands, croplands, and woodlands were higher than those of other LULC types. An area with better consistency can reach more than 50% of the average area of the corresponding land types, but grasslands were mixed with other LULC types in ESRI10 and GLC_FCS30. (4) According to the accuracy assessments, LSV10 data have the highest overall classification accuracy, 79.32%, and the classification accuracy of major land types is also higher than 70%; GLC_FCS30 data have the lowest overall accuracy, 70.14%. Based on these results, LSV10 can more accurately demonstrate LULC than the other four datasets. This study can be used as a reference for selecting land cover data, and it also highlights that the necessary assessments of consistency and accuracy are essential when conducting land use/cover change studies in a specific region.

(Link to Metadata) Circa 2001 land use / land cover (LULC) for the Lake Champlain Basin. The goal in creating this layer was to generate an "improved" version of NLCD 2001 using ancillayr GIS data and Landsat satellite imagery. The Lake Champlain Basin Program (LCBP) is a joint federal-state initiative that helps to monitor and protect Lake Champlain and its contributing watersheds. One area of particular concern is nutrient loading to the Lake, particularly phosphorus (P), from terrestrial and non-lake sources. In order to quantify how much P is entering the lake, it is crucial to have an accurate representation of land use for the Basin. This layer represents an updated digital land use - land cover (LULC) map for the entire Lake Champlain Basin, termed LCB 2001. This updated LULC layer was generated by using an expert system that integrated the 2001 National Land Cover Database (NLCD) with ancillary GIS datasets and circa 2001 Landsat satellite imagery. A primary focus of this expert system was to improve the mapping accuracy of the agriculture LULC class by reducing the confusion with urban open space. An accuracy assessment was carried out by comparing the classification to temporally comparable high resolution imagery. The overall accuracy of LCB 2001 was 88%. The user's accuracy for the urban and agricultural classes, those considered to be the greatest sources of phosphorous, was 84% and 89% respectively. LCB 2001 was produced largely by improving NLCD 2001 using ancillary data. The process of generating LCB 2001 was comprised of three phases: 1) overlay of roads, 2) expert system classification, and 3) assessment and manual correction. Phase 1 was carried out using the aggregate 8-class version of NLCD 2001. The corrected road vector lines were converted to a raster layer with a cell size and alignment matching that of NLCD 2001. The road pixels were incorporated into the NLCD 2001 layer using standard raster overlay procedures in which any pixel in NLCD 2001 that corresponded with a road pixel was reassigned to the urban category. The expert system was employed largely to deal with the accuracy issues surrounding agriculture and urban open land. Edge effects and registration differences between NLCD 2001 and the improved CLU layer made simply overlaying the two an unacceptable solution. To overcome this limitation the expert system was developed and deployed using Definiens Professional software (Definiens AG, Munich, Germany). The expert system took advantage of Definiens Professional's ability to "segment" object polygons from image and thematic raster layers. Image object polygons are groups of pixels with similar spectral and spatial characteristics. Image object polygons allow for the inclusion of rules based on complex topological relationships. Image object polygons for this project were derived from both the spring and fall circa 2001 Landsat satellite scenes, but were constrained to the boundaries of the Improved CLU layer and NLCD 2001. Thus, each object polygon consisted of groups of pixels that were spectrally and spatially similar and share the same attributes with respect to the Improved CLU layer and the NLCD 2001 layer. The expert system first evaluated whether or not the object fell into the confirmed agriculture or urban-open categories based on the Improved CLU layer. If either of these tests proved true then the object was assigned to the corresponding class. If the test failed then the alternate scenarios were evaluated. For objects originally classified as agriculture in NLCD 2001 the object was assigned to the output agriculture class only if the object bordered an object already classified as agriculture (to deal with edge effects and layer alignment issues) or if the object was also in the improved CLU layer's possible agriculture category. This rule ran in an iterative loop to compensate for the fact that once objects were classified as agriculture they would influence other border objects. The rule only stopped executing once all objects were finished changing their class assignment. If the object was not assigned to the output agriculture class at this stage (those classified as agriculture in NLCD 2001, but not in LCB 2001) it was evaluated using a series of spectral and spatial rules to assign it to the output brush or urban-open classes. This set of spectral and spatial rules applied a fuzzy class assignment. The object was considered to be more likely to be brush the darker it was and the further it was from urban areas. The object was considered more likely to be urban if it was near urban areas and brighter. For all other classes the objects adopted the NLCD 2001 class. Following the running of the expert system the output classification was manually compared to the Landsat imagery and any objects that appeared to be misclassified were reassigned. As the goal of the project was to maintain as much consistency with NLCD 2001 as possible the layer was maintained in its original coordinate system - USA Albers Equal Area Conic, USGS Version, NAD83 datum (meters).

High resolution land cover dataset for Baltimore County, MD. Seven land cover classes were mapped: (1) tree canopy, (2) grass/shrub, (3) bare earth, (4) water, (5) buildings, (6) roads, and (7) other paved surfaces. The minimum mapping unit for the delineation of features was set at 8 square meters. The primary sources used to derive this land cover layer were color infrared aerial imagery acquired in 2007 as part of the National Agricultural Imagery Program (NAIP), a normalized Digital Surface Model (nDSM) derived from 2005 LiDAR data, LiDAR intensity data resulting from the 2005 acquisition, building footprints, road polygons, and water polygons.

This land cover dataset is considered current as of August, 2007. Object-based image analysis techniques (OBIA) were employed to extract land cover information using the best available remotely sensed and vector GIS datasets. OBIA systems work by grouping pixels into meaningful objects based on their spectral and spatial properties, while taking into account boundaries imposed by existing vector datasets. Within the OBIA environment a rule-based expert system was designed to effectively mimic the process of manual image analysis by incorporating the elements of image interpretation (color/tone, texture, pattern, location, size, and shape) into the classification process. A series of morphological procedures were employed to insure that the end product is both accurate and cartographically pleasing. No accuracy assessment was conducted, but the dataset was subject to a thorough manual quality control. Over 16,000 corrections were made to the classification.

Land cover and land use in the St. Johns River Water Management District based on 2013-2016 digital orthophotography. This layer is a copy of the 2009 lulc dataset, with updates based on conditions in the 2013-2016 orthophotography. The previous Photointerpretation Key was also revised and updated for this project.Source imagery is county imagery flown primarily by FDOT, at varying dates . The 2014 map used the below imagery:Year Flight Season Counties2013 (Dec 2012 –Mar 2013) Duval2014 (Dec 2013 –Mar 2014) Alachua, Baker, Bradford, Clay, Flagler, Lake, Marion, Nassau, Osceola, Polk, Putnam, St. Johns2015 (Dec 2014 –Mar 2015) Brevard, Indian River, Okeechobee, Seminole, Volusia2016 (Dec 2015 –Mar 2016) OrangeNOTE: Mapping convention made for Marion/Alachua counties. Imagery Dates: Alachua - December 2013 - January 2014; Marion - March 2014 -April 2014. Implications for land cover mapping: We experienced a wetter than average winter, and significant rain fell between Alachua and Marion flight dates. Orange Lake looked very different in areas where the two county imagery datasets overlapped (higher water levels in Marion imagery). We decided on the following mapping convention: always map according to the Marion imagery where it exists.The 2014 data is mapped to the extent of the previous update in 2009 as far as available imagery allowed. There is data to the east of Nassau County (outside our District and Water Basin boundaries) where no imagery was available. The 2014 data extends beyond its previous limit in a few small areas to fulfill an internal staff request.It continues the historical practice of mapping the portion of the Ocklawaha River Basin in Polk County that was transferred to the Southwest Florida Water Management District.NOTE: June 2018 Changes to Dataset:The 2014 dataset was modified in June 2018 based on results from an Accuracy Assessment, These edits changed net acreage counts for some classes as follows:ACREAGE INCREASES: Class 3300 +2.77 ac / Class 4200 +5.12 ac / Class 5200 +7,254.66 ac / Class 5300 +174,082.38 ac / Class 8350 +122.27 acACREAGE DECREASES: Class 1750 -16.64 ac / Class 2110 -100.00 ac / Class 4110 -3.63 ac / Class 4210 -2.00 ac / Class 5100 -181,337.04 / Class 7400 -7.90 ac

This dataset is composed by 85,152 random points throughout the Brazilian territory selected according to a stratified sampling design, based in 127 regular regions and six slope classes (SRTM). Each sample was visually inspected by three independent interpreters, which associated all the land use and land cover (LULC) changes between 1985 and 2018, on a yearly basis, using as reference two Landsat images per year, a MODIS NDVI time series and high resolution images from Google Earth.

This process was guided by a reference labeling protocol which established the follow LULC classes:

• Annual crop: Areas occupied with short to medium-term crops, usually with a vegetative cycle of less than one year, which after harvest needs to be re-planted.
• Aquaculture: Artificial lakes, where aquaculture and/or salt production activities predominate
• Beach and dune (Other): Sandy areas, with bright white color, where there is no vegetation predominance of any kind.
• Forest formation: Vegetation types with predominance of tree species, with continuous canopy formation
• Grassland formation: Grassland formations with predominance of herbaceous stratum
• Mangrove (Other): Dense and Evergreen Forest formations, often flooded by tide and associated with the mangrove coastal ecosystem.
• Mining (Other): Areas where clear signs of extensive mineral extractions are present, shows clear exposure of the soil by the action of heavy machinery. Only regions surrounding the AhkBrasilien (AHK) and the CPRM digital reference data were considered.
• Not observed: Areas blocked by clouds or atmospheric noise, or with absence of ground observation masked out from analysis.
• Other non-forest natural formations: Marshes (with fluvio-marine influence).
• Other non-vegetated area (Other): Non-permeable surface areas (infrastructure, urban expansion or mining) not mapped into their classes
• Pasture: Pasture areas, natural or planted, related with farming activity. In particular in the Pampa and Pantanal biomes part of the area classified as Grassland Formation also includes pasture areas.
• Perennial crop: Areas occupied with crops with a long cycle (more than one year), which allow successive harvests without the need for new crop.
• Rocky outcrop (Other): Naturally exposed rocks without soil cover, often with the partial presence of rupicolous vegetation and high slope.
• Salt flat (Other): "Apicuns" or Salt flats are formations often without tree vegetation, associated to a higher, hypersaline and less flooded area in the mangrove, generally in the transition between this area and the continent.
• Savanna formation: Savanna formations with defined tree and shrub-herbaceous stratum
• Semi-perennial crop: Cultivated areas with sugar cane
• Tree plantation: Planted tree species for commercial use (e.g. Eucalyptus, Pinus and Araucaria)
• Urban infrastructure: Urban areas with predominance of non-vegetated surfaces, including roads, highways and constructions.
• Water: Rivers, lakes, dams, reservoir and other water bodies
• Wetland: Wetlands with fluvial influence or swampy areas

To enable a proper area estimation and accuracy assessment (Stehman, 2014) the dataset is provided with the sampling probability for each sample (brazil_lulc_samples_1985_2018 and brazil_lulc_samples_1985_2018_row_wise) and the sampling weight (brazil_lulc_samples_1985_2018_row_wise), which was adjusted to disregard the "Not observed" class. The number of votes for the associated LULC class (visual interpretation agreement) and an indication if the sample is between two different LULC classes (border flag) are also provided.

A publication describing in detail the methodology used to produce this dataset is under preparation.

This repository contains three datasets developed as supplemental materials for the research study titled "Restoring Wetland Ecosystem Services in Ethiopia’s Ziway-Shalla Sub-Basin: Integrating Advanced Remote Sensing, Quantitative Driver Analysis, and Predictive Land Use Modeling for Targeted Restoration Planning." These datasets support the analysis of land use and land cover (LULC) changes, ecosystem service valuation sensitivity analysis (ESV), and stakeholder-driven identification of LULC change drivers in the Ziway-Shalla Sub-Basin, Ethiopia, to inform targeted wetland restoration strategies. The datasets are:

1. Google Earth Engine Script (ZiwayShalla_LandCover_Classification_1995-2025.js):

A JavaScript script for Google Earth Engine (GEE) that processes Landsat imagery for four different years (October –January/1995-2025) to generate a land cover classification map for the Ziway-Shalla Sub-Basin. The script performs cloud filtering, calculates spectral indices (e.g., NDVI, SAVI, MNDWI), incorporates topographic data (SRTM DEM), and applies a random forest classifier to map eight land cover classes (waterbody, wetland, shrubland, farmland, settlement, forestland, grassland, barrenland). Outputs include classified maps, spectral indices, accuracy metrics (e.g., overall accuracy, kappa coefficient), and variable importance data, exported as GeoTIFFs and CSVs to Google Drive.

2. R Script for ESV and Sensitivity Analysis (ZiwayShalla_ESV_Sensitivity_Analysis.R):

An R script that calculates Ecosystem Service Values (ESV) and conducts sensitivity analysis for LULC types in the Ziway-Shalla Sub-Basin across six years (1995, 2005, 2015, 2025, 2035, 2045). The script uses LULC area data (hectares) and value coefficients (USD/ha/year) to compute ESV, assesses percentage changes in ESV between periods, and evaluates sensitivity by adjusting value coefficients by ±50%. Outputs include ESV values, percentage changes, sensitivity coefficients, and a multi-sheet Excel workbook (LULC_Sensitivity_Results.xlsx) with detailed results.

3. Stakeholder Questionnaire (ZiwayShalla_LULC_Driver_Survey.docx):

A Microsoft Word-formatted questionnaire was designed to collect stakeholder insights on the drivers of LULC change in the Ziway-Shalla Sub-Basin. The survey includes two sections for rating direct (e.g., agricultural expansion, overgrazing) and indirect (e.g., population growth, land tenure insecurity) drivers on a 1–5 importance scale, plus open-ended questions for observations and recommendations. The questionnaire supports quantitative driver analysis and informs restoration planning. It is ready for distribution to stakeholders (e.g., farmers, policymakers) and includes ethical considerations (e.g., confidentiality, informed consent).

Purpose: These datasets enable the reproduction of the study’s analyses, including LULC classification, ESV sensitivity estimation, and driver assessment, to support wetland ecosystem restoration in the Ziway-Shalla Sub-Basin. They are provided to ensure transparency, facilitate reuse by researchers, and comply with open science standards.

Files:

• ZiwayShalla_LandCover_Classification_1995.js: GEE script for land cover classification.
• ZiwayShalla_LandCover_Classification_2005.js: GEE script for land cover classification.
• ZiwayShalla_LandCover_Classification_2015.js: GEE script for land cover classification.
• ZiwayShalla_LandCover_Classification_2025.js: GEE script for land cover classification.
• ZiwayShalla_ESV_Sensitivity_Analysis.R: R script for ESV and sensitivity analysis.
• ZiwayShalla_LULC_Driver_Survey.docx: Word questionnaire for LULC driver assessment.
• README.md: Documentation file with usage instructions for each dataset.

Usage:

• GEE Script: Run in the GEE Code Editor with user-defined ROI and training data to generate land cover maps and metrics.
• R Script: Execute in R/RStudio with updated LULC area data to produce ESV and sensitivity results.
• Questionnaire: Distribute to stakeholders, collect responses, and analyze data (SPSS v20 software). Detailed instructions are provided in the README.md.

Accuracy assessments of classified LULC classes for the years (1993, 2003, 2013 and 2023).

Accuracy assessment of LULC classes for 1988, 1999, 2009 and 2019.

Land use/land cover (LULC) mapping in fragmented landscapes, characterized by multiple and small land uses, is still a challenge. This study aims to evaluate the effectiveness of Synthetic Aperture Radar (SAR) and multispectral optical data in land cover mapping using Google Earth Engine (GEE), a cloud computing platform allowing big geospatial data analysis. The proposed approach combines multi-source satellite imagery for accurate land cover classification in a fragmented municipal territory in Southern Italy over a 5-month vegetative period. Within the GEE platform, a set of Sentinel-1, Sentinel-2, and Landsat 8 data was acquired and processed to generate a land cover map for the 2021 greenness period. A supervised pixel-based classification was performed, using a Random Forest (RF) machine learning algorithm, to classify the imagery and derived spectral indices into eight land cover classes. Classification accuracy was assessed using Overall Accuracy (OA), Producer’s and User’s accuracies (PA, UA), and F-score. McNemar’s test was applied to assess the significance of difference between classification results. The optical integrated datasets in combination with SAR data and derivate indices (NDVI, GNDVI, NDBI, VHVV) produce the most accurate LULC map among those produced (OA: 89.64%), while SAR-only datasets performed the lowest accuracy (OA: 61.30%). The classification process offers several advantages, including widespread spectral information, SAR’s ability to capture almost all-weather, day-and-night imagery, and the computation of vegetation indices in the near infrared spectrum interval, in a short revisit time. The proposed digital techniques for processing multi-temporal satellite data provide useful tools for understanding territorial and environmental dynamics, supporting decision-making in land use planning, agricultural expansion, and environmental management in fragmented landscapes.