RADIOMETRY FOR PREDICTING TALLGRASS PRAIRIE BIOMASS USING REGRESSION AND NEURAL MODELS

Standing forage biomass (SFB) and the percent of standing biomass comp osed of forbs (PCTF) were modeled across the growing season. Samples r epresenting stages of plant maturity from early vegetative to dormant were collected from grazed and ungrazed native tallgrass paddocks usin g a 0.5 x 0.5 m quadrat. Total biomass was measured during all years o f the study (1992-1995). Grass and forb biomass were measured separate ly during 1995. Height of canopy closure also was measured during 1995 . Before clipping, plots were scanned with a multispectral radiometer. Models were prepared using simple regression, multiple regression (MR ), or a commercial neural network (NN) computer program. Potential inp uts to MR and NN models of SFB and PCTF included Julian day of harvest (JD), range site, canopy closure height (CH), incident radiation, spe ctral reflectance values (RFV) at 8 discreet bandwidths, and the norma lized difference vegetation index (NDVI). The NDVI alone accounted for little variability (R-2 = 0.13) in SFB during all years of the study. The optimal MR model for the same data set (SFB = 3.5[JD] -43.7[460 n m RFV] + 1099[NDVI] - 992; R-2 = 0.62) accounted for a greater amount of the variability in SFB. The capacity to describe variation in SFB f or the 1995 data with MR was improved when CH was included as a variab le (R-2 = 0.58 versus 0.78). A NN model accounted for the most variati on in SFB across the entire study (R-2 = 0.76). During 1995, the capab ility of a NN to account for variation in SFB within the training data was similar whether or hot CH was included as an input (R-2 = 0.86); however, prediction of SFB from validation data using the same NN was improved by using CH as an input variable. Little variation in PCTF wa s accounted for by a MR model (R-2 = 0.23); however, a considerably la rger proportion of the variation in PCTF was accounted for when an NN was used (R-2 = 0.59), Seasonal changes in SFB and PCTF were described with an acceptable degree of accuracy by forage reflectance character istics that were adjusted for time of season and canopy complexity, Mo reover, when provided with the same potential inputs, NN predicted SFB and PCTF from validation data more accurately than MR models.