Location: Natural Resource Management ResearchTitle: Prediction of senescent rangeland canopy structural attributes with airborne hyperspectral imagery) Author
Submitted to: GIScience and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/17/2013
Publication Date: 6/14/2013
Publication URL: http://handle.nal.usda.gov/10113/58400
Citation: Phillips, B.L., West, M.S., Saliendra, N.Z., Rundquist, B., Pool, D. 2013. Prediction of senescent rangeland canopy structural attributes with airborne hyperspectral imagery. GIScience and Remote Sensing. 50(2):133-153. Interpretive Summary: Assessment of landscapes is facilitated by synoptic maps of vegetation canopy properties, where ‘snapshots’ of several hundred hectares can be evaluated seasonally or annually with reflectance data. These reflectance data can be modeled to estimate mass of canopy standing crop, canopy nitrogen, bare ground, and photosynthetically-active vegetation (PV), but reflectance models vary with physiography, phenology, and the proportion of non-photosynthetically active vegetation (NPV). Synoptic estimation of these canopy properties could guide rangeland management and biofuel harvest decisions, but approaches need to be tested to constrain sources of variability. Here we present how field data and NASA Jet Propulsion Lab hyperspectral imagery were modeled following removal of grazers in October to represent a 36,000 hectare landscape within the Grand River National Grassland, South Dakota. Field samples collected at random plots were compared with 180 bands of reflectance spectra collected over the same areas. Both Partial Least Squares Regression and re-sampling model selection procedures were employed to analyze the data. Predictions errors are greater for heterogeneous grassland canopies, as compared to homogenous croplands, but models show promise. Results suggest application of topographic, field and reflectance data facilitate a proactive management approach for comprehensive ecosystem assessment aimed at geo-locating low structure post growing-season.
Technical Abstract: Canopy structural and chemical data are needed for senescent, mixed-grass prairie landscapes in autumn, yet models driven by image data are lacking for rangelands dominated by non-photosynthetically active vegetation (NPV). Here, we report how aerial hyperspectral imagery might be modeled to predict canopy attributes post growing-season using two approaches: 1) Application of narrow spectral regions with Vegetation Indices (VIs) and 2) Application of the full spectrum with Partial Least Squares Regression (PLSR). We collected Airborne Visible Infrared Spectrometer (AVIRIS) imagery and field data at 24 random herbaceous plots divided into summit, midslope and toeslope positions (72 sites total). Field data included dry mass for photosynthetically active vegetation (PV), NPV, total standing crop (TSC) and canopy nitrogen (N). We also estimated percent bare ground cover (%BG). We tested established VIs derived from selected regions of the spectrum with a novel re-sampling model selection procedure for each variable. We evaluated all regions of the spectrum with Partial Least Squares Regression (PLSR) for the same dataset. The randomly selected validation dataset (24 of the 72 sites) in the PLSR analyses indicated R2 values for NPV, TSC, %BG, Canopy N, and PV were 0.56, 0.62, 0.58, 0.67, and 0.73, respectively, with prediction errors that were lower than VI models. Analyses of VIs in a re-sampling model selection procedure indicated the short-wave infrared (SWIR) simple ratio ('2128 / '1642) was a key predictor for TSC, NPV and %BG. Overall, the SWIR (from 1260 to 1880 nm) was the spectral region most important for prediction of senescent rangeland canopy attributes. Analyses of the full spectrum using PLSR resulted in slightly lower root-mean-square error of prediction, as compared to VIs, which represent reflectance ratios for specific spectral bands. We conclude that prediction of canopy mass, N content and %BG can be achieved for senescent rangeland landscapes, given hyperspectral imagery and field data.