A Multi-Year Dataset of Annual Inundation Dynamics in Nebraska (2017–2025) Derived from AlphaEarth Foundations Embeddings

# Annual Surface Water Inundation Dataset — Nebraska, USA (2017–2025) **Author:** Aditya Kapoor **Affiliations:** University of Nebraska-Lincoln; Indian Institute of Technology Roorkee **DOI:** https://doi.org/10.5281/zenodo.18741436 **Contact:** Aditya Kapoor --- ## Overview This Zenodo record bundles two complementary data products from the same study area: 1. **Annual surface water inundation classification rasters and inundation-frequency layers** for 901 HydroLAKES surface water bodies across the US state of Nebraska, covering a 9-year period from 2017 to 2025. Generated using AlphaEarth Foundations Embeddings (AFE) — a precomputed, 10-meter-resolution, 64-dimensional annual embedding dataset produced by Google and Google DeepMind — classified within Google Earth Engine (GEE). 2. **(New in v2)** A **validation-data package** containing the in-situ daily water-surface elevation records, elevation-area curves, satellite zonal statistics, and per-reservoir validation tables that underpin an IEEE conference-paper validation of the AFE-derived annual inundation against gauge-derived reference area at five Nebraska reservoirs (Enders, Hugh Butler, Swanson, Harlan County, and Lake McConaughy). --- ## Dataset Contents | File | Size | Description | |---|---|---| | `AEF_RF_2017_2025.zip` | 26.2 MB | Annual inundation rasters from supervised Random Forest classification (2017–2025) | | `AEF_K_MEANS_2017_2025.zip` | 346.7 MB | Annual inundation rasters from unsupervised K-means classification (2017–2025) | | `DYNAMIC_WORLD_DATA_2017_2025.zip` | 40.0 MB | Annual inundation data derived from the Dynamic World product for comparison (2017–2025) | | `INUNDATION_FREQUENCY_RASTERS_RF_2017_2025.rar` | 6.7 MB | Inundation frequency rasters derived from Random Forest classification | | **`AlphaEarth_Nebraska_Validation_Data_v1.zip`** | **0.6 MB** | **(v2 addition)** In-situ daily elevation records, elevation-area curves, satellite zonal statistics, and the per-reservoir validation table for the five reservoirs used in the IEEE conference paper. Contains 14 CSV files and an internal `README.md` data dictionary. | | `ZENODO_README.md` | — | This documentation file | --- ## Data Specifications ### Raster products | Property | Value | |---|---| | Spatial resolution | 10 m | | Temporal coverage | 2017–2025 (annual composites) | | Spatial coverage | State of Nebraska, USA | | Coordinate reference system | EPSG:4326 (WGS 84) | | File format | GeoTIFF | | Pixel values | 1 = Inundated, 0 = Non-inundated, 255 = No data | ### Validation data (inside `AlphaEarth_Nebraska_Validation_Data_v1.zip`) | Property | Value | |---|---| | Spatial coverage | Five validation reservoirs in the Nebraska Great Plains; the summary tables also list Lewis and Clark Lake (excluded from the validation) and, in the two mean-elevation tables, Harry Strunk Lake | | Temporal coverage | 2017-01-01 through 2025-12-31 (daily for in-situ; annual for satellite) | | Sample size | 45 validation reservoir-year pairs (5 reservoirs × 9 years); `annual_area_comparison_2017_2025.csv` has 54 rows, the other nine being Lewis and Clark Lake | | File format | CSV (UTF-8) plus an internal Markdown README | | Reservoirs | Enders, Hugh Butler, Swanson, Harlan County, Lake McConaughy | Refer to the internal `README.md` inside the validation ZIP for the column-by-column data dictionary. --- ## Methodology ### Input Data Annual inundation extents were derived by classifying AFE — a precomputed, analysis-ready embedding dataset that encodes spectral, temporal, and spatial signatures from multiple Earth observation sources at 10 m resolution. The Earth Engine collection is `GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL`. The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind. ### Classification **1. Random Forest (Supervised)** A Random Forest classifier (Breiman, 2001; 150 trees, three variables per split, bag fraction 0.1, at most 100 leaf nodes) was trained on 30,000 labeled points (10,000 per class) for three classes: inundation, vegetation, and built-up area. Training used the 2017 AFE layer, and the classifier was then applied to all years (2017–2025). Training (resubstitution) accuracy was 99.9%. A separate validation set of 7,500 points (2,500 per class), drawn from reference polygons different from the training polygons, was classified on the 2018 layer: overall accuracy was 94.3% and kappa 0.915, with 100% producer's and user's accuracy for the inundation class; all errors were built-up points classified as vegetation. These figures come from the corrected validation set of 16 September 2026; the earlier figure of approximately 99.9% had rested on built-up validation polygons that duplicated the training polygons. The validation points lie inside selected reference polygons (lake interiors for the inundation class), so the figures describe point agreement, not a probability-based estimate of map accuracy. **2. K-means (Unsupervised)** K-means clustering (MacQueen, 1967) with four clusters was applied to the 64-dimensional AFE without labeled training data. Clusters are formed by similarity in the 64-dimensional embedding space. **3. Dynamic World (Benchmark)** Dynamic World (Brown et al., 2022) is a Sentinel-2-based near-real-time land use/land cover product used here as a benchmark for comparison against AFE-based classifications. ### Inundation Frequency Inundation frequency (IF) is the percentage of the nine annual layers (2017–2025) in which a pixel was classified as inundated: the number of inundated years × 100 / 9, stored as an 8-bit integer, which truncates the fraction (for example, eight of nine years is stored as 88). An IF of 100% means the pixel was classified as inundated in all nine annual layers, not that it was continuously wet. A year without a valid classification counts as not inundated. ### Reservoir-scale validation (v2 addition) For each of the five validated reservoirs, daily water-surface elevation records from the U.S. Bureau of Reclamation HydroMet system (Enders, Hugh Butler, Swanson, Harlan County) and the Central Nebraska Public Power and Irrigation District (Lake McConaughy) were averaged to an annual mean elevation. The mean annual elevation was mapped to an in-situ reference surface area by piecewise-linear interpolation on the reservoir-specific elevation-area curve. The curves come from the Bureau of Reclamation 1997 sedimentation surveys of Enders Reservoir and Hugh Butler Lake (Ferrari, 1998a, 1998b), a Bureau of Reclamation area table for Swanson Lake computed in 1984 (U.S. Bureau of Reclamation, 1984), the 2000 area table in the U.S. Army Corps of Engineers regulation manual for Harlan County Lake (U.S. Army Corps of Engineers, 1973), and USGS SIR 2005-5040 for Lake McConaughy (Kress et al., 2005). The AFE Random Forest annual inundation rasters were summarized within the HydroLAKES reference polygon for each reservoir using zonal statistics in EPSG:5070. The resulting 45 reservoir-year pairs are tabulated in `01_validation_summary/annual_area_comparison_2017_2025.csv` inside the validation ZIP, together with nine rows for Lewis and Clark Lake, which was excluded from the validation. --- ## Google Earth Engine Script The GEE analysis script is shared through the Earth Engine Code Editor (opening it requires a Google account with Earth Engine access): https://code.earthengine.google.com/bfaf1067b39ea6b3ef8ad628b673921e Two interactive web applications accompany this dataset: - **Annual classified inundation extents** — https://adityain2003.users.earthengine.app/view/comparison - **Inundation frequency layer** — https://adityain2003.users.earthengine.app/view/inundation-frequency-webapp --- ## Key References - Breiman, L. (2001). Random Forests. *Machine Learning*, 45(1), 5–32. - Brown, C. F., et al. (2022). Dynamic World, near real-time global 10 m land use land cover mapping. *Scientific Data*, 9(1), 251. - Brown, C. F., et al. (2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. *arXiv:2507.22291*. - Ferrari, R. L. (1998a). *Enders Reservoir 1997 sedimentation survey*. Bureau of Reclamation, Technical Service Center, Denver, CO. - Ferrari, R. L. (1998b). *Hugh Butler Lake 1997 sedimentation survey*. Bureau of Reclamation, Technical Service Center, Denver, CO. - Kress, W. H., Sebree, S. K., Littin, G. R., Drain, M. A., & Kling, M. E. (2005). Comparison of preconstruction and 2003 bathymetric and topographic surveys of Lake McConaughy, Nebraska. U.S. Geological Survey Scientific Investigations Report 2005-5040. - MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. In *Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability* (Vol. 1, pp. 281–297). University of California Press. - Messager, M. L., et al. (2016). Estimating the volume and age of water stored in global lakes using a geo-statistical approach. *Nature Communications*, 7, 13603. - U.S. Army Corps of Engineers, Kansas City District (1973). *Republican River Basin Lake Regulation Manual, Volume No. 2: Harlan County Lake, Nebraska* (record copy 2008; Annex I, area table from the 2000 survey). - U.S. Bureau of Reclamation (1984). *Swanson Lake area table in acres*, Frenchman-Cambridge Division (ACAP, computed 10 February 1984). --- ## License and Use This dataset is published under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license. You are free to share and adapt the material for **non-commercial purposes**, provided appropriate credit is given. **Commercial use of the dataset — including use of the data to train commercial machine-learning models or to develop commercial products or services — requires prior written permission from the author.** Please contact Aditya Kapoor. In addition to the formal license terms, researchers planning to use this dataset in a publication, dissertation, or technical report are kindly requested to notify the author prior to manuscript submission. This enables coordination with related ongoing work and avoids inadvertent duplication of analyses. This request supplements, but does not modify, the CC BY-NC 4.0 license under which the data is released. --- ## Citation If you use this dataset, please cite this Zenodo record: > Dataset: https://doi.org/10.5281/zenodo.18741436