Efficient weed segmentation in maize fields: A semi-supervised approach for precision weed management with reduced annotation overhead

Abstract

Listen

Translate article

The rampant growth of weeds poses a significant threat to agricultural productivity, causing substantial damage to crops and hindering overall yield. With the continuous increase in global food demand, effectively addressing the challenges posed by weeds has become an urgent priority. Currently, the application of deep learning methods has proven effective in handling complex classification problems. Nevertheless, their widespread integration into agriculture is impeded by the labor-intensive and time-consuming nature of data annotation. In contrast, the abundance of readily available unannotated data offers an opportunity to develop potent and precise models without incurring the substantial costs associated with data labeling. This study focuses on the development of a semi-supervised deep learning model capable of acquiring semantic segmentation knowledge from both annotated and unannotated images. The research findings indicate that, even with limited data, the performance of the proposed model approaches or exceeds supervised learning approaches that use a larger dataset. The model exhibits resilience in accurately segmenting weeds in high spatial resolution remote sensing images, achieving a performance of 85.5% mIoU on test data, even in scenarios with intensive weed infestation. By significantly reducing annotation costs, our approach contributes to providing a more efficient and cost-effective solution for weed management in agricultural fields.