Winter wheat yield prediction using convolutional neural networks and UAV-based multispectral imagery

An inexpensive and precise crop yield prediction technology is required for facilitating precision agriculture for Asian countries in which small-scale fields are primarily managed. One of the most popular deep learning methods, convolutional neural networks (CNNs), yield better performances for classification problems than other general machine learning techniques. It is necessarily to verify the effectiveness of CNN for crop yield prediction. To do this, UAV-based multispectral imagery was acquired in four growth stages, including the heading, milk, dough, and ripening stages of winter wheat. The effects of growth stage on yield prediction accuracy were assessed. Furthermore, the effects of the combination of different growth stages on accuracy were assessed using multi-temporal CNN model. The prediction accuracies of CNN models were compared with linear regression models based on a typical vegetation index, enhanced vegetation index 2 (EVI2), as a conventional regression algorithm. The CNN model of the heading stage showed the lowest RMSE (0.94 t ha−1) among the four growth stages and outperformed the best linear regression model (RMSE of 1.00 t ha−1). The prediction accuracies of the multi-temporal CNN, and multiple linear regression models based on EVI2 were less than that of the CNN model of the heading stage. These results suggested that the CNN had the potential to improve the accuracy of yield prediction, and the heading stage was suitable data acquisition time for winter wheat in this study. In addition, the combination of growth stage may not improve the accuracy. Further studies with higher resolution multispectral images and integration of weather data are needed to improve the accuracy and robustness of the model and adaptability for different cultivars.