Semi-Supervised Semantic Segmentation for Agricultural Aerial Images

Unmanned Aerial Systems (UAS) have been an essential tool for field scouting, nutrient applications, and farm management. However, assessing the aerial images captured by UAS is labor-intensive, and human assessment can be misleading, introducing bias. Deep learning based image segmentation has been proposed to assist in segmenting different areas of interest in the field, but it usually requires significant pixel-level annotated data. To address this, we propose a semi-supervised learning algorithm, AgSemSeg, to train a robust image segmentation

model with less annotated data. Semi-supervised semantic segmentation aims to predict accurate pixel-level segmentation results via incorporating unlabeled images. Existing methods rely on computing the consistency loss on the output predictions between pseudo-labels and unlabeled images. In AgSemSeg, we exploit the intermediate feature representations rather than only using the output predictions to improve the overall performance of the

model. Specifically, we add a projection layer on the output of the backbone encoder, and inject consistency loss between intermediate feature representations with Sliced-Wasserstein distance. We evaluate AgSemSeg using Agriculture-Vision dataset and outperform the supervised baseline by up to 9.71%. We also evaluate AgSemSeg on benchmark datasets such as PASCAL VOC 2012 and Cityscapes datasets, and it outperforms supervised baselines by up to 24.6% and 7.5% mIoU, respectively. We also perform extensive ablation studies to show that our proposed components are key to the performance improvements of our method.