Skip to main content
Silvilaser 2019

Silvilaser 2019 - Poster Presentations »

Fine-scale prediction of yield in sugarcane: A comparison of UAV-derived LiDAR scans and multispectral imagery

Unmanned Aerial Vehicle (UAV) platforms and associated sensing technologies is a rapidly developing field being extensively used in precision agriculture and farming. Using 3D LiDAR and imaging sensors mounted on small rotorcraft UAVs we can observe fine-scale variations in crops that can help improve the efficiency of fertilizer inputs and maximize yields. In this study we use a combination of LiDAR and multispectral imaging sensors mounted on a DJI M600 Pro quadcopter to map multiple sugarcane nitrogen (N) fertilizer rates field trials in the Wet Tropics region of Australia throughout the 2017-2018 growing season. The LiDAR scans were collected using Velodyne's VLP-16 sensor and processed using Simultaneous Localization and Mapping (SLAM) algorithm, while multispectral imagery was collected using Micasence RedEdge sensor and processed using Pix4D software. From UAV surveys performed every 42 days we generated a time-series of structural and spectral characteristics of the sugarcane crops allowing us to monitor crop growth in terms of height, density and vegetation indices (e.g. normalized difference vegetation index (NDVI)). Furthermore, we created predictive models of sugarcane yields, allowing us to infer the stage at which it is possible to derive reliable at-harvest yield predictions from UAV-derived data at fine scale. The field data for informing UAV-derived data using multiple regression analysis was collected in 2 m transects across crop rows, where sugarcane was cut and weighed for total, leaf and stalk (i.e. yield) biomass. Our preliminary results suggest that LiDAR and multispectral imagery have similar performance in predicting at-harvest yield throughout the growing season, while their combined use increased R2 and decreased RMSE values by 0.08 and 0.5 kg, respectively. The combined use of LiDAR and multispectral imagery resulted in most accurate sugarcane yield prediction models with R2 of 0.63 and RMSE of 4.5 kg at both 100 and 142 days from planting (DFP), which gradually decreased to 0.35 and 6 kg at pre-harvest (i.e. 268 DFP). Interestingly, our predictive models based on N application rate, soil type and NDVI information alone performed about as well as the models based on the combination of LiDAR and multispectral imagery throughout the growing season with highest R2 of 0.6 and RMSE of 4.7 kg at 100 DFP. However, the main advantage of using LiDAR and multispectral imagery in combination is the ability to predict yield in the absence of site-specific information. Our results are of particular interest to nutrient-management programs aiming to deliver soil- and site-specific N fertiliser guidelines for sustainable sugarcane production, as fine-scale yield predictions are feasible when management interventions are still possible (i.e. as soon as at 100 DFP).

Iurii Shendryk
CSIRO
Australia

Jeremy Sofonia
Emesent
Australia

Robert Garrard
CSIRO
Australia

Yannik Rist
CSIRO
Australia

Danielle Skocaj
Sugar Research Australia
Australia

Peter Thorburn
CSIRO
Australia

 


Powered by OpenConf®
Copyright ©2002-2018 Zakon Group LLC