Unsupervised Domain Adaptation for Weed Segmentation Using Greedy Pseudo-Labelling

Abstract

Automatic weed identification based on RGB images with convolutional neural networks (CNN) is a new frontier of precision agriculture. However, the CNN models expect a large volume of labelled data. Their performance deteriorates across different fields due to varied agricultural contexts. To address this, we propose an unsupervised domain adaptation (DA) framework leveraging pseudo-labelling. Our method involves co-training labelled source data with pseudo-labelled target data. We introduce a novel greedy pseudo-labelling strategy to optimize pseudo-label selection, maximizing gains while minimizing overfitting risks. Monitoring overfitting with covariance helps detect fluctuations in class pixel counts during co-training, enhancing target performance. The proposed framework has demonstrated superior performance by evaluation against literature approaches, including the input-level DA methods with Fourier Transform, feature-level with CycleGAN methods and AdaptSegNet, and output-level with self-training. It is tested with the ROSE challenge dataset from different cameras and years with diverse plant stages. Particularly in challenging conditions for plants across different years with varied plant stages, the proposed method outperforms existing literature that struggles to surpass the baseline.