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

# Dataset: Annual Inundation Dynamics in Nebraska (2017-2025) Derived from AlphaEarth Foundation Embeddings ## Author Aditya Kapoor, University of Nebraska-Lincoln ## Description This repository contains annual surface water inundation masks for 901 lakes and wetlands across Nebraska, USA. The data was generated as part of a study evaluating the efficacy of AI-based foundation models for regional hydrological monitoring. The dataset compares supervised and unsupervised classification techniques applied to the AlphaEarth Foundations (AFE) 64-dimensional embeddings, benchmarked against the Sentinel-2 based Dynamic World product. The inundated pixels are valued 1 and the non-inundated are marked with a value of 0. ## Interactive Visualization Tools The following Google Earth Engine (GEE) web applications are available for interactive exploration of this dataset: 1. **[Inundation Comparison App](https://adityain2003.users.earthengine.app/view/comparison)**: Provides a side-by-side visual comparison between AlphaEarth-derived results (RF/K-Means) and the Dynamic World product. 2. **[Inundation Frequency Webapp](https://adityain2003.users.earthengine.app/view/inundation-frequency-webapp)**: Visualizes the long-term Inundation Frequency (IF) layers across Nebraska (2017–2025), identifying permanent vs. ephemeral water bodies. ## Folder Structure & Content ### 1. `AEF_RF_2017_2025/` (Supervised Classification) - **Method:** Random Forest (RF) classifier trained on 30,000 data points (10,000 per class: Inundation, Vegetation, Built-up). - **Data Source:** AlphaEarth Foundations (AFE) Annual Embeddings (10m resolution). - **Files:** Annual binary masks (`1` = Inundated, `0` = Dry, `255` = No Data) for 2017–2025. - **Accuracy:** Validated at 99.9% (Kappa) against field-derived elevation-area relationships. ### 2. `AEF_K_MEANS_2017_2025/` (Unsupervised Classification) - **Method:** K-Means clustering (4 clusters) applied directly to the AFE embeddings without prior training labels. - **Data Source:** AlphaEarth Foundations (AFE) Annual Embeddings (10m resolution). - **Files:** Annual binary masks derived from cluster aggregation. - **Note:** Tends to slightly overpredict inundation in high-moisture transition zones compared to RF. ### 3. `DYNAMIC_WORLD_DATA/` (Benchmark Data) - **Method:** Derived from the Google/WRI Dynamic World V1 product. - **Data Source:** Sentinel-2 Top-of-Atmosphere reflectance. - **Files:** Annual aggregated masks filtered for the "Water" class. - **Comparison:** Used as a standard remote sensing benchmark; our study indicates AFE-based RF provides higher precision in dark/shadow-prone regions. ## Data Specifications - **Spatial Resolution:** 10 meters - **Temporal Coverage:** 2017 - 2025 (Annual Median Composites) - **Geographic Extent:** Nebraska, USA (State-wide) - **Coordinate Reference System:** EPSG:4326 (WGS 84) - **Format:** GeoTIFF ## Usage Notes These masks are intended for researchers working on wetland ecology, water resource management, and geospatial AI.