Predicting willow planting events by conventional machine learning from GPS data: accuracy, generalization ability, and potential for improvement

Willow cultivation is an important activity that can provide greater amounts of cleaner and renewable energy. To support the development of this farming sector, data is required on the spatial distribution of planted plots, as well as on the performance of operations typically required in willow crop management. Unfortunately, this kind of data is largely unavailable, while documenting it is a challenging task. GPS data, which may be feasibly collected during operations, may be a good carrier of information not only to document the spatial location of plots, but also to learn about the frequency of typical events as specific to willow operational management. Based on a GPS dataset characterizing two plots, which was collected during planting operations and labelled manually (8,385 observations), a neural network was used in this study to spatially classify events such as driving (hereafter called D), maneuvering (hereafter called M), planting (hereafter called P), and being stopped (hereafter called S). Three models were trained and validated based on features such as GPS speed (hereafter called model S), GPS speed and leg length (hereafter called model S&L), and GPS speed, leg length, and heading (hereafter called model S&L&H), respectively. Classification performance was found to be impressive, with an overall accuracy of 92.0 (S), 92.1 (S&L), and 93.3% (S&L&H), respectively. The quality of the models was then checked visually using a dataset containing unseen data characterizing two plots of different cardinal orientation, indicating an acceptable generalization ability. The methods described in the paper may be useful when dealing with large datasets and limited resources and expertise in labelling the data manually, as they provide location and event specific data with high accuracy. Improvements in accuracy are possible by integrating the raw data in deep learning, an approach that should be explored further.