Robust UDA for Crop and Weed Segmentation: Multi-Scale Attention and Style-Adaptive Techniques

Abstract

Accurate weed detection and crop mapping are pivotal in precision agriculture. Semantic segmentation methods require labour-intensive pixel labelling. The performance of these methods tends to degrade across different crop fields due to varying agricultural contexts and field conditions. We introduce a novel Unsupervised Domain Adaptation (UDA) framework to overcome these limitations. Our framework operates by style-transforming labelled source domain images to resemble unlabeled target domain images closely. Then, we integrate Enhanced Hybrid Training (EHT) into the framework. EHT combines self-training for generating reliable pseudo-labels of target images and multi-resolution discriminator-based adversarial training to further bridge the domain gap. Unlike previous methods that compromise image resolution, our method effectively combines the strengths of high-resolution image patches for fine segmentation details with low-resolution image patches for capturing long-range context dependencies. The proposed framework demonstrates superior performance compared to the state-of-the-art UDA methods. We evaluate our approach using six public datasets from the ROSE challenge, featuring images from different robots, cameras and years with diverse plant growth stages.