ESTIMATION OF FORAGE PRODUCTION USING MULTISPECTRAL RADIOMETRY K. C. Olson, R. C. Cochran, and G.

Towne

Summary A sixteen-channel multispectral radiometer (MSR; Model 87, Cropscan, Inc.) was used to estimate forage production on native tallgrass prairie pastures during the 1992 - 1995 growing seasons. Approximately 550 plots (1.71 ft 2 ) were hand-clipped at multiple stages of plant maturity ranging from late boot to dormancy in order to measure actual forage production. Plots were scanned prior to clipping using an MSR fitted with filters for measuring reflected sunlight at eight discrete bandwidths (range = 460 to 810 nm). Reflectance characteristics of tallgrass prairie forage were used to predict forage productivity with the aid of a neural network computer program. Forage productivity estimates made using MSR data were found to be closely correlated to those determined by hand clipping (r 2 = 0.70) across all plant growth stages and maturity levels. Determining forage production with multispectral radiometry in combination with a neural network is a reliable alternative to hand clipping and can be accomplished in much less time. Introduction Measurement of pasture forage production is essential in determining range condition and proper stocking rates. Physically measuring forage production by hand clipping is a slow and laborious process and is seldom conducted by livestock producers. Instead, local offices of the Natural Resources Conservation Service conduct periodic surveys of forage productivity at various locations in a Conservation District. These surveys are infrequent and usually will not accurately reflect forage productivity during years when growing conditions are extremely good or poor. This situation can lead to periodic over- or undergrazing which reduce the potential for profit. Development of rapid, cost-effective methods for measurement of forage productivity can alleviate this problem. Multispectral radiometry (MSR) has the potential to predict forage productivity quickly and cheaply. It is based on the principle that every substance absorbs and reflects electromagnetic radiation (i.e. sunlight) in a manner characteristic of its physical and chemical properties. The amount of sunlight reflected by a forage crop is directly proportional to its biomass. The accuracy of MSR for predicting forage production of a particular plot is slightly less than hand-clipping. This is because all reflectance data collected by MSR must first be validated for accuracy against standard hand clipped samples. However, an MSR device can provide more estimates of forage productivity over larger areas than can be accomplished by clipping techniques (approximately 5 seconds for an MSR scan vs. 5 to 10 minutes for a handclipped sample). As a result, the MSR can be used for tasks for which hand-clipping is not well suited. For example, describing the wide variation in productivity across different range sites.

115

Experimental Methods One ungrazed and three grazed tallgrass prairie pastures, located at the Kansas State University Range Research Unit, were used in this study. Three soil types were identified within each pasture before the study began: loamy upland, limestone breaks, and thin claypan. Samples were collected at approximately one week intervals from mid May to mid October. At each sampling date, productivity on 30 to 120 plots was measured with the multispectral radiometer (MSR; Cropscan , model MSR - 87). A total of 551 plots were measured. An equal number of samples were collected on each soil type. After a radiometric measurement was collected on each plot, a 1.71 ft 2 area within the plot was clipped at ground level. Clipped forage was dried and weighed to determine actual forage production. Reflectance data collected with the MSR were used as the inputs for a neural network (Braincel , Promised Land Technologies, Inc.). A neural network is a computer program that simulates the inductive reasoning process in humans. In this case, it was used to predict dry weight of the forage from its reflectance features. The neural network was trained to minimize error on unseen data (Best Net Search option). Production (i.e. dry weight) predicted via the MSR and neural network were compared with the actual weights of clipped samples.

Results and Discussion Twelve categories of information relating to forage characteristics were used as inputs for the neural network (Figure 1). The categories that were most influential in predicting standing dry matter were date of sample collection and reflectance at 460 nm, 560 nm, 610 nm, 660 nm, and 810 nm. The NDVI (normalized difference vegetation index; 810 nm - 610 nm/ 810 nm + 610 run) was also an important factor. The soil type on which individual plots were located and sunlight intensity at the time of sampling (IRR) appeared to be of less importance to predictive ability. Forage productivity, as predicted by the MSR/neural network combination across season and soil type, was able to account for almost 70% of the variation in standing dry weight (Figure 2). The MSR/neural network appeared to predict forage productivity most accurately between 100 g/m2 and 550 g/m 2 . However, the relationship appeared to become weaker when the amount of standing dry matter was above or below these levels. This may have been partially due to the limited number of measurements available for areas with extremely high or low forage productivity. Another factor limiting accuracy of measurements at very high levels of productivity may have been the growth characteristics of the plants themselves. For example, more stem and leaf material is raised above ground level as the plant matures. As a result, leaves and stem material closer to the ground become shaded by the upper parts of the plant and may not reflect sunlight proportional to their surface area. As a final step to test the predictive ability of the MSR/neural network, data from additional plots (approximately 30) that were not included in the training set (i.e., unseen data) were used to predict forage productivity (Figure 3). Once again, the MSR/neural network was able to account for 70% of the variation in forage dry weight, indicating a very strong relationship between forage production and reflectance characteristics of the forage. Validation of the MSR/neural network system for Flint Hills native tallgrass will be completed in the near future. At that time, individuals with similar systems will be able to use the predictive equations developed in this study to estimate the productivity of native tallgrass, eliminating the need for extensive hand-clipping.

116

The MSR/neural network used in this study adequately predicted the weights of clipped forage over a variety of plant growth stages and maturity levels, although the accuracy of prediction appeared to be somewhat less at extremely high or low levels of forage productivity. Radiometers like the one used in this study should prove useful for rapid determination of forage production for stocking rate or rangeland monitoring purposes in the future.

117

Actual Biomass, g/m2

Figure 3. Prediction of Range Forage Standing Biomass by Neural Network from Unseen Data

118

Proceedings sponsored by: ELANCO Animal Health

The Department of Animal Sciences and Industry at Kansas State University greatly appreciates the sponsor of the Range Field Day Proceedings. Sponsorships in no way imply the Departments endorsement of the products and services offered by the sponsors. The Department welcomes inquires from other individuals, associations and firms that may be interested in sponsoring publications.

Cooperative Extension Service, Kansas State University, Manhattan KANSAS

STATE
UNIVERSITY

Issued in furtherance of Cooperative Extension Work, acts of May 8 and June 30, 1914, as amended. Kansas State University, County Extension Councils, Extension Districts, and U.S. Department of Agriculture Cooperating, Richard D. Wootton, Associate Director. All educational programs and materials available without discrimination on the basis of race, color, national origin, sex, age, or disability.