Unsupervised Dust Removal Method for Tillage Quality Monitoring in Challenging Agricultural Conditions

Autonomous agricultural machinery faces problems such as challenging weather conditions or moving objects. In addition, autonomous tractors must monitor the tillage quality to guarantee the same working quality the whole time. Sensor systems for tillage quality monitoring must cope with the rough working conditions of agricultural machinery. Especially in dry conditions, stirred-up dust heavily affects sensor systems. This problem is addressed in this work that investigates removing dust to improve tillage quality monitoring. First, a camera was mounted to the rear of a disc harrow to record images during tillage. Since the images showed larger parts of the surrounding environment, the images were cropped to the working area, which is important for tillage quality monitoring. The cropped images formed the dataset that was used for further investigations. We used CycleGAN, an unsupervised learning method, to remove the dust from the images. Finally, the dust removal capabilities were evaluated based on the texture features from the gray-level co-occurrence matrix. Our results showed that thin dust could be removed while preserving the field structure. However, thick dust led to artifacts in the image. The dataset and further information are available at https://agriscapes-dataset.com/.