AI-driven onboard hyperspectral data processing for precision farming and environmental monitoring

Hyperspectral imaging (HSI) provides detailed spectral measurements across hundreds of wavelength channels, enabling advanced Earth-observation applications such as land-cover monitoring, atmospheric analysis, and environmental hazard detection. However, the high dimensionality of hyperspectral data introduces substantial computational, memory, and bandwidth demands that hinder real-time or onboard processing on resource-constrained platforms. This thesis investigates dimensionality reduction as a strategy for enabling deployable artificial intelligence pipelines for hyperspectral Earth observation. The research focuses primarily on PRISMA satellite hyperspectral imagery, which provides 239 spectral channels across the visible, near-infrared, and short-wave infrared regions, of which 234 bands are typically used for analysis after removal of unstable channels. Depending on the application scenario, both the full spectral representation and a filtered 189-band configuration are considered to mitigate sensor artefacts and low signal-to-noise spectral regions. Two dimensionality reduction approaches are investigated: linear compression using Principal Component Analysis (PCA) and nonlinear representation learning using autoencoder (AE) networks. Rather than applying PCA as a cascaded compression stage, PCA is used to estimate intrinsic spectral dimensionality and guide the design of a compact AE latent space. The resulting representations are evaluated across multiple Earth-observation tasks, including land-cover classification using PRISMA scenes over urban regions in Italy, aerosol optical thickness (AOT) estimation using radiative-transfer-based simulated spectra, and wildfire detection using PRISMA observations acquired during major bushfire events in Australia. Additional validation is conducted using UAV datasets for agricultural monitoring, including RGB weed detection and hyperspectral crop mapping benchmarks. This multi-task evaluation framework enables systematic analysis of how dimensionality reduction influences predictive performance across classification, regression, and segmentation scenarios. Results show that hyperspectral data contain substantial inter-band redundancy, allowing compact latent representations to retain task-relevant information while significantly reducing computational complexity. PCA demonstrates strong performance for physics-driven regression tasks such as AOT estimation, whereas nonlinear autoencoder representations provide advantages in heterogeneous classification and detection scenarios such as wildfire monitoring. Finally, the proposed pipelines are deployed and profiled on representative edge-AI platforms, analysing latency, memory footprint, and computational scaling under different spectral representations. The results indicate that reducing spectral dimensionality substantially lowers inference cost while maintaining competitive predictive performance, supporting the feasibility of real-time hyperspectral AI under embedded constraints. Overall, this thesis demonstrates that carefully designed spectral representations can bridge the gap between hyperspectral information richness and practical deployment requirements, establishing dimensionality reduction as a key enabler for scalable hyperspectral Earth-observation analytics.