Sequential forward selection and support vector regression in comparison to LASSO regression for spring wheat yield prediction based on UAV imagery

Traditional plant breeding based on selection for grain yield is time-consuming and costly; therefore, new innovative methods are in high demand to reduce costs and accelerate genetic gains. Remote sensing-based platforms such as unmanned aerial vehicles (UAV) show promise to predict different traits including grain yield. Attention is currently being devoted to machine learning methods in order to extract the most meaningful information from the massive amounts of data generated by UAV images. These methods have shown a promising capability to come up with nonlinearity and explore patterns beyond the human ability. This study investigates the application of two different machine learning based regressor methods to predict wheat grain yield using extracted vegetation indices from UAV images. The goal of the study was to investigate the strength of Support Vector Regression (SVR) in combination with Sequential Forward Selection (SFS) for grain yield prediction and compare the results with LASSO regressor with an internal feature selector. Models were tested on grain yield data from 600 plots of spring wheat planted in South-Eastern Norway in 2018. Five spectral bands along with three different vegetation indices; the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and MERIS Terrestrial Chlorophyll Index (MTCI) were extracted from multispectral images at three dates between heading and maturity of the plants. These features for each field trial plot at each date were used as input data for the SVR model. The best model hyperparameters were estimated using grid search. Based on feature selection results from both methods, NDVI showed the highest prediction ability for grain yield at all dates and its explanatory power increased toward maturity, while adding MTCI and EVI at earlier stages of grain filling improved model performance. Combined models based on all indices and dates explained up to 90% of the variation in grain yield on the test set. Inclusion of individual bands added collinearity to the models and did not improve the predictions. Although both regression methods showed a good capability for grain yield prediction, LASSO regressor proved to be more affordable and economical in terms of time.