Predictive performance of a machine learning model for plot-level dry matter yield estimation using drone derived multispectral indices.

Accurate and non-destructive herbage mass estimation has significant potential for both commercial farm management and research applications. High resolution drone imagery can potentially provide an avenue to estimate herbage yield through the collection of multispectral data. However, optimising the use of multispectral vegetation indices remains a key challenge. This study evaluated the performance of an XGBoost gradient boosting framework incorporating multiple multispectral vegetation indices to predict plot-scale dry matter yield (kg DM/ha) in perennial ryegrass (Lolium perenne L.).

Imagery was captured across plot trials using a DJI Mavic 3M multispectral sensor. The following spectral vegetation indices were calculated from the imagery; Normalised Difference Vegetation Index (NDVI), Green Normalised Difference Vegetation Index (GNDVI), and Normalised Difference Vegetation Index Red-Edge (NDVIre), alongside actual measured plot level dry matter yield (kg DM/ha). An XGBoost regression model was trained using 5-fold cross-validation with early stopping. Feature importances were calculated using fractional gain.

The XGBoost model predicted plot dry matter yield with an R² of 0.48, and a Root Mean Square Error  of 395.87 kg DM/ha. The trial accounted for significant baseline variation, while NDVI (Gain: 0.133) and NDVIre (Gain: 0.115) provided the highest relative predictive contribution among spectral features, outperforming GNDVI (Gain: 0.102).

Combining multispectral indices with gradient boosting provides a viable workflow for dry matter yield estimation, offering an efficient methodology for high-throughput assessment.