The Presence and Background Learning (PBL) framework has demonstrated a robust performance for Semi-Supervised Learning (SSL) techniques in remote sensing, particularly in case of the scarce labelled data. However, it’s a conventional reliance on random sampling to form the background set introduces a critical vulnerability: random selection may inadvertently include pixels from the target class, thereby corrupting the background distribution and degrading the model adjustment. In our paper, we propose a novel modified PBL algorithm that replaces the random background sampling with a confidence-driven pseudo-label selection strategy. Precisely, our naïve PBL model is first trained on a small set of labelled positive samples and a preliminary set of unlabelled samples; its predicted probability scores are then used to filter the full unlabelled pool; only samples with a low presence probability, those unlikely to belong to the target class, are retained as the background. This produces a cleaner, more semantically coherent background distribution, thereby improving the model training. Applied to a water body extraction from a Sentinel-2 imagery over the ecologically complex of the El Kala National Park (Algeria), our method consistently outperforms the naïve PBL baseline across all evaluation metrics, including the Kappa coefficient and F1-score.
|
