A case study on canola (Brassica napus L.) potential yield prediction using remote sensing imagery and advanced data analytics

Abstract

Listen

Translate article

Canola (Brassica napus L.) yield prediction using a combined application of small unmanned aerial system (sUAS) and vegetation indices (VIs) have gained significant attention. In recent years, major studies have demonstrated the potential of developing new VIs for predicting canola yield. However, such indices may perform optimally on the specific farms for which they have been designed and may fail to generalize to a new agronomic scenario. Therefore, this study aims to conduct a comprehensive analysis on the application of existing VIs that could be used to predict potential canola yield on a commercial farm, thereby eliminating the need to develop new indices from scratch. In this research study, over 27 VIs were extracted from two sUAS imagery captured during peak flowering and seed development stages of canola. The extracted features were fed to four conventional machine learning (ML) classifiers with appropriate hyperparameter tuning approaches. Additionally, to perform a comparative test with neural network-based deep learning (DL) architectures, a convolutional neural network -1-dimensional (CNN-1D_Canola) model was developed to train data points and predict canola yield. These models were trained on the yield maps interpolated using three approaches, based on the ground truth yield data points obtained from a harvester. Results suggest that peak flowering is the best stage to predict canola yield. Additionally, a combination of kriging-based yield maps with three best features, canola ratio index (CRI), canola index (CI), and structure intensive vegetation index (SIPI) indices, trained using SVM (R2=0.68), MLP (R2=0.7), and CNN_1D Canola (R2=0.66) have the potential to predict canola yield based on spectral image-based features. This study highlights the potential of predicting canola yield for a commercial farm using a combined application of sUAS imagery and VIs. The promising performance of all the models coupled with a comprehensive hyperparameter tuning approaches suggests its applicability in predicting canola yield in real field conditions.