Spatial Feature Engineering Dataset for Forest Aboveground Biomass Estimation Using Landsat Imagery

Study Area:The dataset covers forested regions in Oregon, Washington, Idaho, and eastern Montana, characterized by diverse climatic conditions due to orographic effects. The forests in the Coast Range and western slopes of the Cascades, with high precipitation (800-3000 mm annually), contrast with the drier forests in Idaho and Montana, which receive over 400 mm annually. The dataset includes highly productive Douglas-fir and western hemlock forests, with aboveground biomass (AGB) densities exceeding 1200 Mg ha⁻¹, as well as fire-adapted lodgepole and ponderosa pine forests in the rainshadow regions.

LiDAR AGB Estimates:The dataset includes 176 lidar-derived AGB maps from 2002 to 2016, covering various regions in Oregon, Washington, Idaho, and Montana. A Random Forest (RF) model was used to estimate AGB at a 30m² resolution, utilizing lidar height features, DEM features, and climate data. Non-forested areas and buildings were masked using binary forest cover maps from the LCMS dataset and the Microsoft Building Footprints dataset.

Reference Dataset:A composited AGB map, derived from the 176 lidar maps, was created to develop Landsat-based AGB models, covering 9,361,622 ha of forested land. The AGB layer was stratified into 30 bins, and training, development, and testing sets were constructed for model validation. The dataset includes 7500 test samples and 300,000 training and development samples, with a 500m buffer around test set locations to prevent spatial autocorrelation.

Landsat Satellite Imagery:Landsat imagery from 1990 to 2022 was utilized, with three time series derived: all scenes, scenes from May to November, and annual medoid composites. The imagery was processed using the Google Earth Engine (GEE) platform, focusing on periods of maximum phenological activity.

Feature Engineering:Extensive feature engineering was performed, generating spectral, spatial, temporal, and topographic features from Landsat imagery and DEM data. Features were extracted over the reference AGB map's domain, synchronized with the lidar acquisition dates.

LandTrendr Fitted Imagery: Spectral features were derived from LandTrendr-fitted imagery, smoothing variations in the time series.

LandTrendr Disturbance and Recovery Features: Temporal features were derived from LandTrendr models, characterizing disturbance and recovery events.

CCDC Disturbance and Recovery Features: CCDC algorithm-derived features characterized disturbances and recovery using harmonic models.

Buffer Features: Local variations were captured using buffer statistics around each pixel.

GLCM Features: GLCM texture features summarized the joint distribution of gray-tone values.

Edge Detectors: Various edge detection operators captured spatial derivatives and edges.

Morphological Operations: Morphological features were derived using multi-channel image processing techniques.

Neighborhood Vectorization: Direct vectorization of satellite measurements in pixel neighborhoods.

Neighborhood Similarity: Similarity features characterized the relationship between pixel neighborhoods and their centroids.

Topographic Features: Topography was characterized using elevation, slope, aspect embeddings, and hillshade layers from the NED DEM.

This comprehensive dataset enables robust analysis of AGB models and their performance across diverse forested landscapes in the Pacific Northwest