Lāna‘i Landcover Mapping Input Geopackages

This dataset provides high-resolution, species-specific land cover maps for the Hawaiian island of Lāna'i based on 2020 WorldView-2 satellite imagery. Machine learning models were trained on extensive ground control polygons and points. The land cover maps capture the distribution and diversity of vegetation with high accuracy to support conservation planning and monitoring. This data release consists of two child items, one containing the field and expert collected ground control data used to train our models, and another consisting of resulting land cover maps for the island of Lāna‘i. The research effort that generated these input data, and products are carefully described in the associated manuscript Berio Fortini et al. 2024. Full citation is listed in the larger work section of this XML file. Inputs included in this page include: Ground control polygons used for model training and evaluation (ground_control_polygons.gpkg): This dataset consists of refined vegetation polygons digitized across the island of Lāna‘i representing the 15 land cover classes of interest. High-resolution aerial imagery and extensive field experience were used to iteratively collect and improve the polygons through expert review and interpretation. The polygons were divided into a 250m grid overlaying the island to balance sample size and spatial resolution while reducing spatial autocorrelation, resulting in 1,754 smaller polygons. These polygon data served as the primary dataset used to train, validate, and evaluate the classification models through cross-validation. An iterative collection process aimed to achieve satisfactory model accuracy across all classes prior to final model selection and island-wide mapping. Ground control points used for independent pixel-level model validation (ground_control_points.gpkg): This dataset consists of 313 points distributed across the 15 vegetation classes on the island of Lāna‘i. The points were randomly generated from the final species-specific land cover classification map and stratified by class to ensure representation across all classes. The dataset provides species-specific land cover labels for the 313 points, with the spatial location corresponding to the pixel coordinate location on the 2m resolution land cover map. Comparing modeled class assignments to these expert-validated classes enables an independent accuracy assessment supplemental to the polygon-based cross-validation accuracy evaluation.

[
  "010:12"
]

[
  "geospatial"
]