Muhammad, Hassan (2025) EXPLAINABLE DEEP LEARNING FOR PLANKTON CLASSIFICATION: BRIDGING THE GAP BETWEEN AI AND MARINE BIOLOGY. [Tesi di dottorato]

Documento PDF EXPLAINABLEDEEPLEARNINGFORPLANKTONCLASSIFICATIONBRIDGINGTHEGAPBETWEENAIANDMARINEBIOLOGY.pdf Visibile a [TBR] Amministratori dell'archivio Download (33MB) | Richiedi una copia

Abstract

Plankton taxonomic classification is a major application enhancing research in marine ecology, ocean monitoring, and biodiversity studies. Traditional taxonomic methods take a long time and are error-prone, while deep learning methods are afflicted with problems such as limited labeled data, imbalanced classes, computational ineffectiveness, and lack of interpretability. This PhD thesis overcomes the above limitations through two complementary research contributions. In the first work, a hybrid deep learning architecture based on transfer learning was introduced by combining InceptionResNetV2 and a novel DeepPlanktonNet learned from scratch. Feature fusion was applied to merge high-level and domain-specialized representations, and the Whale Optimization Algorithm (WOA) reduced redundancy for computational efficiency. For interpretability, Local Interpretable Model-agnostic Explanations (LIME) provided post-hoc interpretability of predictions. The approach achieved state-of-the-art accuracy and interpretability, thus making automatic ecological monitoring more efficient. The second contribution introduced a multimodal learning framework that incorporated deep features of InceptionResNetV2 together with handcrafted features (SIFT-BoVW), benefitting from the complementary strengths of deep and handcrafted descriptors. LIME was also introduced to enable explainability for handcrafted and deep features. The model significantly improved classification accuracy over baseline Convolutional Neural Network (CNNs) and made interpretable decision support possible for marine scientists. Combined, these contributions show that hybrid architectures, fusion of features, optimization, and explainability can be used to improve both accuracy and transparency in plankton classification. The results of this thesis offer concrete foundations for future intelligent ecological monitoring systems to aid in supporting marine conservation, climate research, and sustainable management of resources.

Downloads

Actions (login required)

Modifica documento