Exploiting centimetre resolution of drone-mounted sensors for estimating mid-late season above ground biomass in rice

Abstract

Above ground biomass (AGB) is an important indicator of rice for improving agronomic management efficiency and yield monitoring in crops. In particular, rice AGB during the mid (reproductive) and late (ripening) stages are responsible for the panicles per given area, the number of spikelets or grains per panicle, the percentage of filled kernels and grains; and the weight of each grain. Consequently, proper monitoring of rice AGB, particularly during the mid to late growth stages, are important for accurate estimation of rice yield. To this end, monitoring AGB at centimetre scale has become implementable by using sensors onboard Unmanned Aerial Vehicles (UAVs) or drones. The RGB sensors capable of generating plant height estimations from digital surface models provide a viable option for monitoring rice AGB. The advancement in miniature Multi-Spectral Imager (MSI) sensors capable of generating vegetation indices (VIs) and texture metrics (TM) also provides the opportunity to ascertain the capability of the sensor to estimate rice AGB, particularly during the growth stages. The study compares the potential and relative merits of using drone-mounted consumer-grade RGB imagery and/ or scientific-grade multispectral imagery for estimating rice mid-late season above ground biomass. Plant height estimates generated from digital surface model derived from the RGB sensor were compared with in-situ measurements of biomass using a simple linear regression (SLM) model. On the other hand, VIs, TM and their combination were accessed using the Random Forest model for estimating rice AGB. We also accessed the combination of both sensors for estimating rice AGB. Results testing model quality statistically showed plant height (R2 = 0.72; RMSE = 1.07 t/ha; MAE = 0.93 t/ha) estimates from the RGB camera performed better than VIs (R2 = 0.59; RMSE = 1.31 t/ha; MAE = 1.06 t/ha), TM (R2 = 0.43; RMSE = 1.58 t/ha; MAE = 1.22 t/ha) and the combination of VIs and TM when estimating rice AGB at the mid to late growing stages. When combining plant height and VIs from both cameras to estimate AGB, results suggest that the combination using random forest models improve the estimation of rice AGB. The combination of TM, VIs and Plant Height (PH) estimates produced the most statistically accurate estimates (R2 = 0.74; RMSE = 1.02 t/ha; MAE = 0.82 t/ha). Our findings suggest that the Plant height estimates from the RGB sensor produce a more accurate estimation of AGB compared to the MSI camera. However, the most accurate estimations are seen when both sensors are combined to estimate rice AGB at the mid to late growth stage.