Authors: K.J. Han¹, M.W. Alison², X. Meng³, B. Awasthi³, K.C. Manisha³, and V. Sehgal¹

• ¹ School of Plant, Environmental, and Soil Sciences, LSU AgCenter, Baton Rouge, LA
• ² Macon Ridge Research Station, LSU AgCenter, Winnsboro, LA
• ³ Department of Geography and Anthropology, LSU, Baton Rouge, LA

Estimating pasture yield traditionally involves labor-intensive and time-consuming methods like multiple spot clippings, drying biomass samples, and weighing. While simplified methods such as rising plate meters or grazing sticks exist, they often have significant error potential due to the limited representativeness of entire pasture.

A viable alternative could be using a prediction model developed with multispectral data from a UAV-mounted camera. Field trial data from two Louisiana State University Agricultural Center Research Stations, Southeast and Macon Ridge, were regressed against twelve light indices calculated from multispectral data to predict forage yield using a random forest algorithm.

Before each harvest, light reflection data were collected from annual ryegrass and oats in the test plots. To account for potential differences between ryegrass and oats, two modeling approaches were attempted: one using only ryegrass data and another using a composite of ryegrass and oats data. Model training and validation were conducted by shuffling the training and test datasets at a 3:1 ratio 200 times.

The observed yield data demonstrated a moderate skew towards higher yields across three harvests. Among the 12 light indices, Red Edges and NDVI (Normalized Difference Vegetation Index) were the most significant predictors for yield estimation. The average RMSE (root mean square error) was 363 lbs per acre for the ryegrass-only model and 411 lbs per acre for the composite model. The average correlation between prediction and observation was 0.97 in both models, indicating their robustness.

These yield prediction models appear applicable for estimating overseeded cool-season annual forage biomass yield at moderate biomass levels. By incorporating more datasets, such as multispectral and yield data from warm-season grass pastures, the modeling can be expanded to estimate yields for a wider variety of forage types. This approach is anticipated to not only enhance accuracy but also significantly reduce the labor and time required compared to traditional methods, making it a promising tool for stakeholders in agriculture.

Contact: khan@agcenter.lsu.edu