Table 1.
Generalized ecological state classes found within the focal study region of the USDA ARS Jornada Experimental Range.
Fig 1.
Conceptual workflow illustrating the hyper-temporal remote sensing modeling framework.
Modeling steps include the compilation of Landsat TM 5 NDVI layers from 1984–2011, the decomposition and extraction of the seasonal and trend components of the time series for each pixel stack, and the integration of the modified time series and sampled field data into a support vector machine classification model to produce ecological site class prediction maps.
Fig 2.
(a) Study region and site location in southwestern NM, USA; (b) boundary of the Jornada Long-term Ecological Research site with delineated location of focal study area; and (c) focal study area underlain by 1-m National Agriculture Imagery Program (NAIP) imagery and overlain with 176 sampling locations (red circles).
Fig 3.
gSSURGO maps of ecological sites.
(a) dominant ecological sites, and (b) secondary ecological sites aggregated by dominant condition.
Fig 4.
Expert delineated and SVM predicted maps of ecological sites.
(a) expert delineated ecological sites and (b) support vector machine predicted ecological sites.
Table 2.
Model performance for SVM using leave-group-out cross validation and comparison to expert delineated and gSSURGO classified ecological sites.
Table 3.
Error matrices for ecological site classes predicted using support vector machine (SVM) classification, expert delineation and gSSURGO classification.
Fig 5.
Examples of unfiltered NDVI hyper-temporal spectra from selected sampling points for the Clayey and Deep Sand ecological site classes.
Clayey sites represent a sequence from least disturbed (a) to increasing levels of disturbance (b, c). Similarly, Deep Sand sites represent a sequence from least disturbed (d) to increasing levels of disturbance (e, f).
Fig 6.
Barplots illustrating model misclassification by ecological site and state.
(a) the percentage of misclassified points within each ecological state code; (b) the sample size distribution of all points (black) and all misclassified points (grey) broken out by state code; and (c) all misclassified points that had a shrub woodland state (state code 6) broken out by ecological site. Also, the percentage of misclassifications within each ecological site class that had a state code of 6 (c) is listed at the base of each bar.
Fig 7.
Results on the effect of ecological state on the misclassification of the Loamy ecological site.
The frequency distribution of ecological states is shown for (a) all correctly classified Loamy sites and their corresponding spectral signature (mean +/- 1SD); (b) all misclassified Loamy sites and their corresponding spectral signature; (c) all Loamy sites misclassified as Sandy and their corresponding spectral signature; (d) all Loamy sites misclassified as Deep Sand and their corresponding spectral signature; and (e) all Loamy sites misclassified as Clayey and their corresponding spectral signature.
Fig 8.
Temporal patterns in SPI and NDVI, and their relationship to SVM variable importance.
(a) standardized precipitation index (SPI) calculated across a range of time scales illustrating general climatic states (e.g., wet, dry normal) at coarse temporal scales (~35 months) and event driven responses at short time-scales (~0–6 months); (b) mean +/- 1 standard deviation of NDVI calculated from Landsat pixels at all 176 sample locations; and (c) tri-panel illustrating relationship between SVM variable importance (VI) for ecological site prediction model (middle panel) and both short-term SPI (top panel) and NDVI standard deviation (bottom panel). See methods section for description of interpreting SPI time-scales.