This thesis proposes a Deep Learning model based on a Transformer Encoder architecture for the simultaneous estimation of eight physico-chemical soil properties from Visible and Near-Infrared (Vis-NIR) spectra. Trained on the Open Soil Spectral Library (OSSL), the architecture integrates both spectral and contextual tokens (encompassing instrumental, geographical, stratigraphic, and radiometric metadata) to disambiguate the spectral signal. The framework employs a multi-task learning approach, utilizing an adaptive loss function weighted by homoscedastic uncertainty to effectively handle samples with partial annotations. The evaluation protocol implements a rigorous spatial group-based partitioning strategy (grouped by geographical site) to preclude data leakage, complemented by the Monte Carlo Dropout method for sample-specific predictive uncertainty estimation. Finally, the high computational efficiency of the model, characterized by a static memory footprint of merely 3.37 MB, renders the system highly suitable for deployment on edge computing devices, thereby enabling in-situ inference.
Il presente lavoro di tesi propone un modello di Deep Learning basato su Transformer Encoder per la stima simultanea di otto proprietà chimico-fisiche del suolo a partire da spettri Vis-NIR. Addestrata sull'Open Soil Spectral Library (OSSL) , l'architettura integra token spettrali e contestuali (strumentali, geografici, stratigrafici e radiometrici) per disambiguare il segnale spettrale. Il framework sfrutta un approccio multi-task learning con loss adattiva basata sull'incertezza omoschedastica per gestire sample con annotazioni parziali. La validazione impiega una rigorosa strategia di partizionamento per raggruppamento spaziale (sito geografico) al fine di prevenire il data leakage , affiancata dal metodo Monte Carlo Dropout per la stima dell'incertezza predittiva sample-specific. L'elevata efficienza computazionale (3.37 MB di memoria statica) rende il sistema idoneo al deployment su dispositivi edge per inferenze in-situ.
A TRANSFORMER-BASED APPROACH FOR ESTIMATING FARMING SOIL PROPERTIES VIA VIS-NIR SPECTROSCOPY
DE FEUDIS, FRANCESCO
2025/2026
Abstract
This thesis proposes a Deep Learning model based on a Transformer Encoder architecture for the simultaneous estimation of eight physico-chemical soil properties from Visible and Near-Infrared (Vis-NIR) spectra. Trained on the Open Soil Spectral Library (OSSL), the architecture integrates both spectral and contextual tokens (encompassing instrumental, geographical, stratigraphic, and radiometric metadata) to disambiguate the spectral signal. The framework employs a multi-task learning approach, utilizing an adaptive loss function weighted by homoscedastic uncertainty to effectively handle samples with partial annotations. The evaluation protocol implements a rigorous spatial group-based partitioning strategy (grouped by geographical site) to preclude data leakage, complemented by the Monte Carlo Dropout method for sample-specific predictive uncertainty estimation. Finally, the high computational efficiency of the model, characterized by a static memory footprint of merely 3.37 MB, renders the system highly suitable for deployment on edge computing devices, thereby enabling in-situ inference.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/110808