Mapping floristic gradients of forest composition using an ordination-regression approach with landsat OLI and terrain data in the Central Hardwoods region

Mapping Forest Composition with Landsat Time Series: An Evaluation of Seasonal Composites and Harmonic Regression

“…Specifically, plot-wise values from each feature set were sampled according to an area-weighted mean of pixel values (as reference plots were not typically centered on pixels and may include more than one pixel) intersecting a 30-m resolution square window (native Landsat resolution) centered on each reference sample location ( Figure 2 displays the proximity and spatial relationship between the sampling grains of the three reference datasets and individual Landsat pixels). This window size was carefully selected as an adequate solution for stabilizing coherence in spatial attributes (i.e., sampling grain), Landsat image resolution, and geolocational precision, among the imagery and reference data at the finest grain size possible [34]. While coarsening the sampling grain, including multi-pixel kernels, has been shown to improve model accuracy in other applications [56], we were largely constrained by rapid transitions in species composition indicative of the dissected topography of the region (e.g., many ridges are <30 m wide).…”

“…Only records with trees of ≥8 cm DBH were used in further analysis from the SILVAH dataset. The second dataset included a dense vegetation inventory (hereafter vegetation plots; n = 699 plots) collected using a fixed-radius (11.3 m;~400 m 2 ) protocol, established following a spatially-stratified cluster sampling design among several forest units in the center of the study area [34]. Trees and other live stems ≥8 cm DBH were identified and DBH recorded within each plot.…”

“…Forest compositional modeling has been shown to be dramatically improved in our study area with the incorporation of ancillary datasets joined with multispectral imagery [34]. Therefore, we obtained a 10-m digital elevation model (DEM) derived from USGS 7.5 minute quadrangle contours from the Ohio Environmental Agency Division of Emergency and Remedial Response.…”

“…Classification, by labeling of vegetation type to map unit, has remained the predominant approach to landscape-scale vegetation mapping [1,[27][28][29]. Gradient modeling has emerged as a means to reconcile conventional mapping practice with a reality that ecological communities display variation in stand transitional patterns beyond that of categorical maps [30][31][32][33][34][35][36]. Specifically, the aim of gradient modeling is the development of a non-classificatory approach that strives to represent pixels as collections of individual species without reference to any specific class membership, while preserving continuous properties of floristic change, as a result of intergrading stands of co-occurring species that gradually transition from one community to the next.…”

“…Specifically, the aim of gradient modeling is the development of a non-classificatory approach that strives to represent pixels as collections of individual species without reference to any specific class membership, while preserving continuous properties of floristic change, as a result of intergrading stands of co-occurring species that gradually transition from one community to the next. Various gradient-based approaches, such as those based on canonical correspondence analysis [30] or flexible ordination-regression techniques [34,35,37], are similar concerning the transfer of detailed, plot-level species data to ordination models, in which extracted ordination values among the principal axes are utilized as response inputs to approximate compositional attributes of individual forest stands. Embodied in the overall gradient-based approach is the community-continua concept of community ecology, which combines both (i) aspects of traditional classification, and (ii) the envisioning of communities as representing collections of overlapping yet independent species distributions [35,38].…”