Optimizing Landsat time series length for regional mapping of lidar-derived forest structure

Abstract

The value of combining Landsat time series and airborne laser scanning (ALS) data to produce regional maps of forest structure has been well documented. However, studies are often performed over single study areas or forest types, preventing a robust assessment of the approaches that produce the most accurate estimates. Here, we use Landsat time series data to estimate forest attributes across six Canadian study sites, which vary by forest type, productivity, management regime, and disturbance history, with the goal of investigating which spectral indices and time series lengths yield the most accurate estimates of forest attributes across a range of conditions. We use estimates of stand height, basal area, and stem volume derived from ALS data as calibration and validation data, and develop random forest models to estimate forest structure with Landsat time series data and topographic variables at each site. Landsat time series predictors, which were derived from annual gap-free image composites, included the median, interquartile range, and Theil Sen slope of vegetation indices through time. To investigate the optimal time series length for predictor variables, time series length was varied from 1 to 33 years. Across all six sites, increasing the time series length led to improved estimation accuracy, however the optimal time series length was not consistent across sites. Specifically, model accuracies plateaued at a time series length of ~15 years for two sites (R2 = 0.67–0.74), while the accuracies continued to increase until the maximum time series length was reached (24–29 years) for the remaining four sites (R2 = 0.45–0.70). Spectral indices that relied on shortwave infrared bands (Tasseled Cap Wetness and Normalized Burn Ratio) were frequently the most important spectral indices. Adding Landsat-derived disturbance variables (time since last disturbance, type of disturbance) did not meaningfully improve model results; however, this finding was largely due to the fact that most recently disturbed stands did not have predictions of forest attributes from ALS, so disturbed sites were poorly represented in the models. As model accuracies varied regionally and no optimal time series length was found, we provide an approach that can be utilized to determine the optimal time series length on a case by case basis, allowing users to extrapolate estimates of forest attributes both spatially and temporally using multispectral time series data.