Artificial intelligence for biodiversity: understanding the living world through machine vision

The loss of biodiversity represents a major environmental emergency of our time, with profound repercussions on ecosystems and the services they provide. The global recognition of the need for quantitative, reliable, and spatially explicit indicators for ecological monitoring is now globally recognized. However, traditional field-based methods are constrained by high costs, limited spatial coverage, and poor temporal replicability. In this context, the integration of proximal remote sensing through unmanned aerial vehicles (UAVs) and artificial intelligence (AI) techniques offers new opportunities to more efficiently, standardized, and scalably analyze and understand biodiversity. The thesis Artificial Intelligence for Biodiversity: Understanding the Living World Through Machine Vision explores the combined use of UAV imagery and machine learning models to estimate key ecological parameters related to vegetation and pollinators. The research is articulated into three main studies: (I) the ability of RGB UAV imagery and classical machine learning algorithms to estimate floral cover and infer bee community metrics is evaluated; (II) a generalizable model capable of maintaining robust performance across different grassland ecosystems; (III) the third applies convolutional neural networks (U-Net) to multispectral imagery for the semantic segmentation of vegetation and the extraction of reproducible landscape metrics. Overall, this study demonstrates how artificial intelligence (AI) can bridge the gap between ecological detail and spatial coverage, enabling the automated derivation of ecological indicators at operational scales. The results of this study contribute to the definition of a reproducible and transparent approach to biodiversity monitoring, fostering the integration of advanced computational techniques into conservation and environmental management programs.