A novel semi-supervised approach for semantic segmentation of aerial remote sensing images under limited ground-truth availability

Conventional semantic segmentation techniques rely heavily on the availability of substantial ground-truth data. However, this prerequisite often proves infeasible in real-world scenarios, particularly with the labeling complexities inherent in remote sensing images. In this manuscript, a semi-supervised approach has been investigated towards semantic segmentation of remotely sensed images by addressing the challenge of limited availability of ground-truth information. For this purpose, a hybrid integration of a standard semantic segmentation model and an adversarial model has been proposed under semi-supervised setting. The former predict the masks for the unlabelled images when fine-tuned with the available labelled training images (however limited they may be); whereas the latter aids the reconstruction of original input images from the predicted soft(masks) through an adversarial mechanism. This reconstruction, further validated through a reconstruction score, assist in the identification of ‘most-confident’ image-mask pairs to be strategically integrated into the training set. The contribution ultimately is to utilise the unannotated images to meaningfully augment the limited training set to obtain an enhanced one. The proposed technique showcases a significant improvement, with an 11–34% enhancement over existing approaches in terms of mean intersection over union, precision, and F1-score across both the minifrance and dense labeling remote sensing dataset datasets.