Get reliable rangeland science

Assessments of herbage biomass and forage quality using field hyperspectral (HS) sensing provide valuable support to farmers in making precise forage management decisions. The field HS data, which includes measurements of canopy reflectance in the visible and near-infrared wavelength range (400-2350 nm), has been extensively studied in grassland assessment research. Partial least squares (PLS) regression has been widely adopted as a standard calibration method for estimating herbage biomass (BM) and determining forage quality parameters, such as crude protein (CP) and neutral detergent fiber (NDF) concentrations. In this study, a one-dimensional convolutional neural network (1D-CNN) model was developed as a non-destructive and rapid method for evaluating forage composition. The relationships between HS measurements taken on the ground and forage components obtained through harvest and chemical analysis at ground level were analyzed. The dataset in the orchard grass-dominated meadow field consisted of 200 samples from seven fields in three regions of Hokkaido, Japan, surveyed prior to the first grass harvest in May/June 2023. Overall, the 1D-CNN models showed better predictive accuracies for most parameters (BM, CP, and NDF) than standard PLS regressions. The 1D-CNN model demonstrated a good predictive accuracy (R² = 0.950) for BM, but less accurate predictions for concentrations of CP (R² = 0.650) and NDF (R² = 0.506). However, when the content in percentage was converted to standing mass (g/m²), high predictive accuracies in CP mass (R² = 0.814) and NDF mass (R² = 0.837) were achieved. These results are expected to contribute to the advancement of forage management by enabling rapid and accurate evaluation of forage components.