Using Landsat-Based Phenology Metrics, Terrain Variables, and Machine Learning for Mapping and Probabilistic Prediction of Forest Community Types in West Virginia

Abstract

This study investigates the mapping of forest community types for the entire state of West Virginia, USA using Global Land Analysis and Discovery (GLAD) Phenology Metrics analysis ready data (ARD) derived from the Landsat time series and digital terrain variables derived from a digital terrain model (DTM). Both classifications and probabilistic predictions were made using random forest (RF) machine learning (ML) and training data derived from ground plots provided by the West Virginia Natural Heritage Program (WVNHP). The primary goal of this study is to explore the use of globally consistent ARD data for operational forest type mapping over a large spatial extent. Mean overall accuracy calculated from 50 model replicates for differentiating seven forest community types using only variables selected from the 348 GLAD Phenology Metrics used in the study resulted in an overall accuracy (OA) of 53.36% (map-level image classification efficacy (MICE) = 0.42). Accuracy increased to a mean OA of 73.0% (MICE = 0.62) when the Oak/Hickory and Oak/Pine classes were combined to an Oak Dominant class. Once selected terrain variables were added to the model, the mean OA for differentiating the seven forest types increased to 61.58% (MICE = 0.52). Our results highlight the benefits of combining spectral data and terrain variables and also the enhancement of the product’s usefulness when probabilistic prediction are provided alongside a hard classification. The GLAD Phenology Metrics did not provide an accuracy comparable to those obtained using harmonic regression coefficients; however, they generally outperformed models trained using only summer or fall seasonal medians and performed comparably to spring medians. We suggest further exploration of the GLAD Phenology Metrics as input for other spatial predictive mapping and modeling tasks.