Dosi, Andrea (2026) A Deep Learning Approach to the Analysis of Complex Datacubes. [Tesi di dottorato]

Documento PDF dosi_andrea_38.pdf Download (12MB)

Abstract

Modern scientific imaging increasingly relies on the analysis of high-dimensional data acquired through complex and often indirect measurement processes. Hyperspectral Earth observation and radio astronomy are two representative examples of this trend: both generate structured datacubes whose interpretation is far from trivial and requires models able to combine local correlations with global context, often under significant uncertainty. Against this background, this thesis focuses on the design of efficient deep learning architectures for hyperspectral image analysis and frames these developments within the broader perspective of inverse problems in scientific imaging. The first part of the thesis introduces AMBER, an advanced SegFormer-based architecture designed for semantic segmentation of hyperspectral datacubes. This chapter reports on the corresponding paper already published for Neural Computing and Applications, Springer Nature[2]. One of AMBER’s main features is its ability to operate directly on hyperspectral datacubes, without preprocessing with dimensionality reduction methods (e.g., SVD). The architecture (the transformer) was designed to jointly encode spectral and spatial information: a 3D convolution combined with multi-head self-attention. The thesis then extends the preprint "Less is More: AMBER-AFNO – a New Benchmark for Lightweight 3D Medical Image Segmentation". In this lightweight AMBER variant, Multi-Head Self-Attention is replaced by Adaptive Fourier Neural Operators[4], and the decoder is suited for 3D segmentation. Despite significantly reducing model complexity -far fewer parameters compared with the SOTA models in this field- this design retains the ability to capture global dependencies, making it particularly attractive for high-dimensional inputs. The work subsequently moves toward the foundation model paradigm by pretraining the AMBER-AFNO encoder in a self-supervised manner (using DINO) on the large-scale SpectralEarth hyperspectral dataset, and then evaluates the semantic segmentation performance on the DESIS-CDL dataset. In this setting, AMBER-AFNO emerges as a hyperspectral foundation model, capable of learning representations that transfer across sensors, geographic regions, and downstream semantic segmentation tasks. Finally, these outcomes are put into perspective with respect to the inverse problems in astrophysics. Although the physics of the two problems is obviously distinct, hyperspectral imaging and astrophysical observation have in common that they are measurement processes governed by an operator which gives rise to high-dimensional data cubes. From this perspective, this thesis highlights some computational patterns and suggests that there is value in architectural solutions that aim at capturing global structure with a reasonable computational cost.

Downloads

Actions (login required)

Modifica documento