Assistant Professor University of Wisconsin-Madison
Disclosure: Disclosure information not submitted.
Water is a critical resource for agricultural production. Currently over 40% of irrigation water is withdrawn from groundwater and increasing demand has been linked directly to reduced aquifers. Potato is a shallow rooted crop that often grown on soils with low to medium water-holding capacities, and it is extremely sensitive to water stress. Because of its high economic value, growers may tend to apply more-than-enough amounts of water and nutrients as an “insurance” to minimize potato production risks. However, potato yield, size, grade, and internal and external quality can all be negatively impacted by either over- or under-irrigation. Improved irrigation management can be achieved by use of new technologies that monitor the real-time crop status.
Traditional destructive sampling approach is tedious and laborious when recording yield and other potato quality data. Recently, unmanned aerial vehicles (UAVs) have gained significant attention in precision agriculture, because of their capability to quickly gather and analyze data that aid in decision making. In addition, hyperspectral imagery consists of much more spectral bands, usually hundreds of them, arranged in a narrower bandwidth, thus providing more detailed spectral information which is of high importance for analyzing the chemical and physical properties of the plants. Therefore, in this work, we propose to develop machine learning models to predict potato yield by using the UAV-based hyperspectral imagery. The calibrated hyperspectral reflectance data were used as inputs of the machine learning models. Different machine learning models include linear regression, support vector regression, random forest, gradient boosting tree, and deep neural networks were explored, and the results were compared with the ground truth yield. The best prediction performance was obtained when the time series hyperspectral image features were combined, and the prediction accuracies were achieved more than 60% for the yield, and 70% for the tuber set.