ALBERT

Description: We propose a Bayesian approach for semantic segmentation of crops and weeds. Farmers often manage weeds by applying herbicides to the entire field, which has negative environmental and financial impacts. Site-specific weed management (SSWM) considers the variability in the field and localizes the treatment. The prerequisite for automated SSWM is accurate detection of weeds. Moreover, to integrate a method into a real-world setting, the model should be able to make informed decisions to avoid potential mistakes and consequent losses. Existing methods are deterministic and they cannot go beyond assigning a class label to the unseen input based on the data they were trained with. The main idea of our approach is to quantify prediction uncertainty, while making class predictions. Our method achieves competitive performance in an established dataset for weed segmentation. Moreover, through accurate uncertainty quantification, our method is able to detect cases and areas which it is the most uncertain about. This information is beneficial, if not necessary, while making decisions with real-world implications to avoid unwanted consequences. In this work, we show that an end-to-end trainable Bayesian segmentation network can be successfully deployed for the weed segmentation task. In the future it could be integrated into real weeding systems to contribute to better informed decisions and more reliable automated systems.