Remote Sensing (Sep 2021)

Improving Accuracy of Herbage Yield Predictions in Perennial Ryegrass with UAV-Based Structural and Spectral Data Fusion and Machine Learning

  • Joanna Pranga,
  • Irene Borra-Serrano,
  • Jonas Aper,
  • Tom De Swaef,
  • An Ghesquiere,
  • Paul Quataert,
  • Isabel Roldán-Ruiz,
  • Ivan A. Janssens,
  • Greet Ruysschaert,
  • Peter Lootens

DOI
https://doi.org/10.3390/rs13173459
Journal volume & issue
Vol. 13, no. 17
p. 3459

Abstract

Read online

High-throughput field phenotyping using close remote sensing platforms and sensors for non-destructive assessment of plant traits can support the objective evaluation of yield predictions of large breeding trials. The main objective of this study was to examine the potential of unmanned aerial vehicle (UAV)-based structural and spectral features and their combination in herbage yield predictions across diploid and tetraploid varieties and breeding populations of perennial ryegrass (Lolium perenne L.). Canopy structural (i.e., canopy height) and spectral (i.e., vegetation indices) information were derived from data gathered with two sensors: a consumer-grade RGB and a 10-band multispectral (MS) camera system, which were compared in the analysis. A total of 468 field plots comprising 115 diploid and 112 tetraploid varieties and populations were considered in this study. A modelling framework established to predict dry matter yield (DMY), was used to test three machine learning algorithms, including Partial Least Squares Regression (PLSR), Random Forest (RF), and Support Vector Machines (SVM). The results of the nested cross-validation revealed: (a) the fusion of structural and spectral features achieved better DMY estimates as compared to models fitted with structural or spectral data only, irrespective of the sensor, ploidy level or machine learning algorithm applied; (b) models built with MS-based predictor variables, despite their lower spatial resolution, slightly outperformed the RGB-based models, as lower mean relative root mean square error (rRMSE) values were delivered; and (c) on average, the RF technique reported the best model performances among tested algorithms, regardless of the dataset used. The approach introduced in this study can provide accurate yield estimates (up to an RMSE = 308 kg ha−1) and useful information for breeders and practical farm-scale applications.

Keywords