Integration of Semi-Supervision and Active Learning for Effective Semantic Segmentation of Remote Sensing Images With Sparse Training Set

In remote sensing, semantic segmentation is frequently constrained by the intricate complexities of the images and the scarcity of annotated data. To address this, a hybrid approach has been proposed that leverages semi-supervised learning and active learning to enhance the segmentation performance. The approach begins by fine-tuning a conventional semantic segmentation model to predict (soft)masks for unlabeled images. Images in the likes of the original unlabelled images are reconstructed from the (soft)masks realized through an adversarial model. The reliable (most-confident) image-mask pairs undergo weighed integration into the training set. To further improve the model’s robustness, active learning is used to incorporate 15% and 10% most challenging (least-confident) predictions, ensuring the model adapts to difficult cases. This method achieves significant gains in segmentation accuracy, with 4-33% improvement in mean intersection over union (mIoU) and overall accuracy (OA) on Dense Labeling Remote Sensing Dataset (DLRSD) and DeepGlobe benchmarks, demonstrating the effectiveness of combining semi-supervised and active learning in accurate and low-cost remote sensing applications.