The Application of an Unmanned Aerial System and Machine Learning Techniques for Red Clover-Grass Mixture Yield Estimation under Variety Performance Trials

Remote Sensing

Karli Sepp¹  Ming-Der Yang²  Miguel Villoslada Peciña³  Raul Sampaio de Lima³  Kalev Sepp³  Kai-Yun Li³  Are Selge³  Janar Raet³  Ants Vain³  NiallG. Burnside⁴ 

[1] Agricultural Research Center, 4/6 Teaduse St., 75501 Saku, Estonia;Department of Civil Engineering, and Innovation and Development Center of Sustainable Agriculture, National Chung Hsing University, Taichung 402, Taiwan;Institute of Agriculture and Environmental Sciences, Estonian University of Life Sciences, Kreutzwaldi 5, EE-51006 Tartu, Estonia;School of Environment & Technology, University of Brighton, Lewes Road, Brighton BN2 4JG, UK;

关键词: unmanned aerial system;    red clover;    random forest;    support vector regression;    artificial neural network;    tillage;   

DOI  :  10.3390/rs13101994

来源: DOAJ

【 摘 要 】

A significant trend has developed with the recent growing interest in the estimation of aboveground biomass of vegetation in legume-supported systems in perennial or semi-natural grasslands to meet the demands of sustainable and precise agriculture. Unmanned aerial systems (UAS) are a powerful tool when it comes to supporting farm-scale phenotyping trials. In this study, we explored the variation of the red clover-grass mixture dry matter (DM) yields between temporal periods (one- and two-year cultivated), farming operations [soil tillage methods (STM), cultivation methods (CM), manure application (MA)] using three machine learning (ML) techniques [random forest regression (RFR), support vector regression (SVR), and artificial neural network (ANN)] and six multispectral vegetation indices (VIs) to predict DM yields. The ML evaluation results showed the best performance for ANN in the 11-day before harvest category (R² = 0.90, NRMSE = 0.12), followed by RFR (R² = 0.90 NRMSE = 0.15), and SVR (R² = 0.86, NRMSE = 0.16), which was furthermore supported by the leave-one-out cross-validation pre-analysis. In terms of VI performance, green normalized difference vegetation index (GNDVI), green difference vegetation index (GDVI), as well as modified simple ratio (MSR) performed better as predictors in ANN and RFR. However, the prediction ability of models was being influenced by farming operations. The stratified sampling, based on STM, had a better model performance than CM and MA. It is proposed that drone data collection was suggested to be optimum in this study, closer to the harvest date, but not later than the ageing stage.

【 授权许可】

Unknown