Unsupervised Domain Adaptation applied to Agriculture and Urban Forests using Transformers, GANs and Diffusion Models

Abstract

Deep learning architectures, such as ConvNets, represented an impressive advance in the field of machine learning and are continually breaking records in numerous areas of artificial inteligence, such as image recognition. Nevertheless, the success of these architectures depends on a large amount of labeled data. The annotation of training data is a costly process often performed manually. In agricultural and urban forest problems, differences in image acquisition conditions, such as the height of capture, different sensors, soil conditions, crop stages, and lighting, often necessitate retraining models as new acquisitions are made. In this context, domain adaptation presents itself as a promising alternative to deal with this situation. Domain adaptation consists of adapting the knowledge learned from a source domain to apply it in a target domain, different but related to the original. The aim of this work is to use the domain adaptation approach to find solutions that address problems requiring large amounts of annotated data. Our focus consist of problems related to agriculture and urban forests, using recent architectures used in the unsupervised domain adaptation, such as Generative Adversarial Networks, Vision Transformers, and Diffusion Models. In this work, we propose an approach to address the problem of detecting crop rows and gaps using dilation to generate approximate segmentation maps from annotated one-pixel-wide lines. This method speeds up the pixel labeling process and reduces the line detection problem to semantic segmentation. We use the transformer-based model DAFormer to evaluate the ability to transfer the knowledge learned from source datasets to target datasets. Additionally, we propose a method for segmenting trees that integrates domain adaptation with image-to-image translation models and super-resolution networks to enhance the quality of low-resolution aerial images. Our method also aims to address the challenge of limited labeled data by employing data augmentation to generate additional high-resolution training samples from the existing labeled data, thereby improving model performance and reducing the need for costly labeling processes.