The health of marine ecosystems is crucial for planetary well-being, yet monitoring these dynamic and often remote environments remains a pressing challenge, further exacerbated by the impacts of climate change. Rising ocean temperatures, acidification, and habitat degradation are altering biodiversity distributions, increasing the need for rapid advancements in monitoring techniques. While current data collection techniques yield valuable ecological insights, they are frequently hindered by logistical constraints, high operational costs, and limited spatiotemporal coverage, underscoring the need for innovative, interdisciplinary approaches. The PEARL framework addresses this gap by integrating probabilistic ecosystems assessment and reinforcement learning (RL) to develop adaptive, explainable tools for biodiversity monitoring in the context of autonomous underwater vehicles (AUVs). At the core of PEARL lies HexaWorld, a reinforcement learning environment structured on hexagonal tiling, designed to optimize AUV navigation in complex marine landscapes. By incorporating partial observability and a multi-objective reward function, HexaWorld enables AUVs to strategically explore biodiversity hotspots while efficiently balancing energy consumption and travel distance, mitigating the inefficiencies inherent in traditional grid-based approaches. Complementing this exploration strategy, Generalized Mixture Models (GMMs), incorporating both positive and negative weights, leverage environmental covariates collected during navigation to model biodiversity distributions. Unlike their positive-weighted counterparts, these models offer a more compact and accurate representation of complex, multimodal patterns, improving the identification of biodiversity hotspots and their relationship with underlying environmental drivers. By embedding GMMs within the RL framework, PEARL translates optimal exploration policies into explainable ecological insights, enhancing the transparency of data-driven decisions and supporting more informed conservation and restoration strategies.

Probabilistic Ecosystems Assessment with Reinforcement Learning. Shaping Explainable AI for Marine Biodiversity Monitoring

Lombardi, Giulia
2025

Abstract

The health of marine ecosystems is crucial for planetary well-being, yet monitoring these dynamic and often remote environments remains a pressing challenge, further exacerbated by the impacts of climate change. Rising ocean temperatures, acidification, and habitat degradation are altering biodiversity distributions, increasing the need for rapid advancements in monitoring techniques. While current data collection techniques yield valuable ecological insights, they are frequently hindered by logistical constraints, high operational costs, and limited spatiotemporal coverage, underscoring the need for innovative, interdisciplinary approaches. The PEARL framework addresses this gap by integrating probabilistic ecosystems assessment and reinforcement learning (RL) to develop adaptive, explainable tools for biodiversity monitoring in the context of autonomous underwater vehicles (AUVs). At the core of PEARL lies HexaWorld, a reinforcement learning environment structured on hexagonal tiling, designed to optimize AUV navigation in complex marine landscapes. By incorporating partial observability and a multi-objective reward function, HexaWorld enables AUVs to strategically explore biodiversity hotspots while efficiently balancing energy consumption and travel distance, mitigating the inefficiencies inherent in traditional grid-based approaches. Complementing this exploration strategy, Generalized Mixture Models (GMMs), incorporating both positive and negative weights, leverage environmental covariates collected during navigation to model biodiversity distributions. Unlike their positive-weighted counterparts, these models offer a more compact and accurate representation of complex, multimodal patterns, improving the identification of biodiversity hotspots and their relationship with underlying environmental drivers. By embedding GMMs within the RL framework, PEARL translates optimal exploration policies into explainable ecological insights, enhancing the transparency of data-driven decisions and supporting more informed conservation and restoration strategies.
25-lug-2025
Inglese
Bianchi, Luigi Amedeo
Università degli studi di Trento
TRENTO
222
File in questo prodotto:
File Dimensione Formato  
Giulia-Lombardi-PhD-thesis-PEARL.pdf

accesso aperto

Dimensione 28.28 MB
Formato Adobe PDF
28.28 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/218561
Il codice NBN di questa tesi è URN:NBN:IT:UNITN-218561