Yield monitoring systems are widely used commercially in grain crops to map yields at a scale of a few meters. However, such high-resolution yield monitoring and mapping for hay and forage crops has not been commercialized. Most commercial hay yield monitoring systems only obtain the weight of individual bales, making it difficult to map and understand the spatial variability in hay yield. This study investigated the feasibility of an unmanned aerial vehicle (UAV)-based remote sensing system for the estimation and mapping of hay yield by machine learning models. Data were obtained during harvest of a 35-ha hay field with mixture of red clover and timothy grass in June of 2021. A multispectral camera consisting of five bands (red, blue, green, near infrared, and red-edge) attached to a UAV was used to acquire images at a flight height of 20 m above ground level, resulting in a ground sampling distance of 18 mm/pixel. For calibration, 110 ground truth hay yield measurements were collected from 1 m2 quadrats. Sixty image features, including vegetation indices and texture features were calculated from the images, and were used to estimate the hay mass yield. For yield estimation, a simple random forest machine learning model was compared with a complex deep learning model, the convolutional neural network (CNN). Using recursive feature elimination, we selected explanatory features for the input dataset of the CNN model, which included a rectified linear unit activation function and a batch normalization process for increasing training speed and creating deeper layers. Additionally, dataset augmentation was applied due to the relatively small size of the CNN calibration dataset. The results of this research provide information to aid in selection of an appropriate analysis method for hay estimation using UAV imagery. In future research, the models developed here will be applied to whole-field imagery for creating hay yield maps.
