DeepVeg: Deep Learning Model for Segmentation of Weed, Canola, and Canola Flea Beetle Damage

Farmers around the world face the challenge of growing more food for the increasing world population. On top of that, external threats such as pests (weeds and insects) pose a threat to crop production and it is necessary to take early steps to reduce the risk. This paper presents semantic segmentation of canola field images collected under natural conditions. The dataset contains four unbalanced classes; background, crop, weeds, and damages in the crop. The damages to the crop leaves are small round shaped and share the same texture and colour as whitish stones from the background. We propose, DeepVeg, a deep learning segmentation model that focuses on the smallest (damage) class without affecting other classes to solve the class imbalance issue. Early stage canola field image dataset is utilized for training and testing the proposed model. Evaluation results show that the proposed method outperforms the benchmark deep learning models and effectively addresses the weed and damaged canola plants segmentation problem. The DeepVeg model demonstrates a superior mean intersection over union score greater than 0.76 and $accuracy$ above 0.97 for four class segmentation. The model also shows robustness in detecting unlabelled, newly grown weeds and canola and is also able to distinguish the similar rounded structured canola plant and weed with small amounts of data for model training, which is suitable for early stage damage and weed segmentation.