<p>Calcareous algae are essential components of marine ecosystems, playing a key role in the biogeochemical cycling of carbonates and contributing to the biodiversity of the South Atlantic. However, offshore oil and gas extraction, which involves subsea equipment and pipelines, has increasingly threatened this environment. To address these impacts, a Brazilian oil and gas company monitored calcareous algae along the coast using a Remotely Operated Vehicle, supported by a deep learning classification model for species identification from images. A major challenge in this task was that the training dataset used contained noisy samples, i.e., mislabeled instances, which can compromise the robustness of deep learning models. To address this issue, state-of-the-art (SOTA) methods adopt the small-loss strategy, assuming that correctly labeled samples yield lower training losses, while samples with higher losses are considered noisy and are removed from the training process. Building upon this approach, the Retrieving Discarded Samples (RDS) framework was recently introduced as a mechanism to recover samples initially excluded during training. In this study, we demonstrate that the performance of the RDS framework is sensitive to the quality of pseudo-labels assigned in the process. To improve this aspect, we introduce RDS-Contrastive Learning, a novel variant that incorporates self-supervised learning to enhance pseudo-label accuracy and overall model performance. We evaluate our model on four different benchmark datasets, achieving up to a 3% improvement in F1-score over existing SOTA methods. In the specific case of calcareous algae monitoring, our model achieved a 1.6% improvement in F1-score compared to other SOTA approaches.</p>