Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DeepSalt: Bridging Laboratory and Satellite Spectra through Domain Adaptation and Knowledge Distillation for Large-Scale Soil Salinity Estimation

Rupasree Dey, Abdul Matin, Everett Lewark, Tanjim Bin Faruk, Andrei Bachinin, Sam Leuthold, M. Francesca Cotrufo, Shrideep Pallickara, Sangmi Lee Pallickara

TL;DR

DeepSalt tackles the challenge of estimating soil salinity at regional to global scales by bridging laboratory FTIR spectroscopy and satellite hyperspectral sensing. It introduces a Transformer-based pipeline that (i) pretrains a lab-focused teacher on FTIR data, (ii) uses a Spectral Adaptation Unit to align lab and satellite spectra in a shared latent space, and (iii) distills spectral knowledge into a multimodal student that fuses adapted spectral features with ancillary covariates for robust salinity prediction. The approach delivers state-of-the-art accuracy (e.g., MAE $0.25$, $R^2=0.87$, RMSE $0.72$) and demonstrates strong generalization to unseen regions, validating cross-domain spectral transfer for scalable soil health monitoring. By enabling accurate, data-efficient salinity mapping without exhaustive ground sampling, DeepSalt offers practical impact for precision agriculture and land management under climate stress, and sets a path for applying similar lab-to-satellite transfer to other soil properties.

Abstract

Soil salinization poses a significant threat to both ecosystems and agriculture because it limits plants' ability to absorb water and, in doing so, reduces crop productivity. This phenomenon alters the soil's spectral properties, creating a measurable relationship between salinity and light reflectance that enables remote monitoring. While laboratory spectroscopy provides precise measurements, its reliance on in-situ sampling limits scalability to regional or global levels. Conversely, hyperspectral satellite imagery enables wide-area observation but lacks the fine-grained interpretability of laboratory instruments. To bridge this gap, we introduce DeepSalt, a deep-learning-based spectral transfer framework that leverages knowledge distillation and a novel Spectral Adaptation Unit to transfer high-resolution spectral insights from laboratory-based spectroscopy to satellite-based hyperspectral sensing. Our approach eliminates the need for extensive ground sampling while enabling accurate, large-scale salinity estimation, as demonstrated through comprehensive empirical benchmarks. DeepSalt achieves significant performance gains over methods without explicit domain adaptation, underscoring the impact of the proposed Spectral Adaptation Unit and the knowledge distillation strategy. The model also effectively generalized to unseen geographic regions, explaining a substantial portion of the salinity variance.

DeepSalt: Bridging Laboratory and Satellite Spectra through Domain Adaptation and Knowledge Distillation for Large-Scale Soil Salinity Estimation

TL;DR

DeepSalt tackles the challenge of estimating soil salinity at regional to global scales by bridging laboratory FTIR spectroscopy and satellite hyperspectral sensing. It introduces a Transformer-based pipeline that (i) pretrains a lab-focused teacher on FTIR data, (ii) uses a Spectral Adaptation Unit to align lab and satellite spectra in a shared latent space, and (iii) distills spectral knowledge into a multimodal student that fuses adapted spectral features with ancillary covariates for robust salinity prediction. The approach delivers state-of-the-art accuracy (e.g., MAE , , RMSE ) and demonstrates strong generalization to unseen regions, validating cross-domain spectral transfer for scalable soil health monitoring. By enabling accurate, data-efficient salinity mapping without exhaustive ground sampling, DeepSalt offers practical impact for precision agriculture and land management under climate stress, and sets a path for applying similar lab-to-satellite transfer to other soil properties.

Abstract

Soil salinization poses a significant threat to both ecosystems and agriculture because it limits plants' ability to absorb water and, in doing so, reduces crop productivity. This phenomenon alters the soil's spectral properties, creating a measurable relationship between salinity and light reflectance that enables remote monitoring. While laboratory spectroscopy provides precise measurements, its reliance on in-situ sampling limits scalability to regional or global levels. Conversely, hyperspectral satellite imagery enables wide-area observation but lacks the fine-grained interpretability of laboratory instruments. To bridge this gap, we introduce DeepSalt, a deep-learning-based spectral transfer framework that leverages knowledge distillation and a novel Spectral Adaptation Unit to transfer high-resolution spectral insights from laboratory-based spectroscopy to satellite-based hyperspectral sensing. Our approach eliminates the need for extensive ground sampling while enabling accurate, large-scale salinity estimation, as demonstrated through comprehensive empirical benchmarks. DeepSalt achieves significant performance gains over methods without explicit domain adaptation, underscoring the impact of the proposed Spectral Adaptation Unit and the knowledge distillation strategy. The model also effectively generalized to unseen geographic regions, explaining a substantial portion of the salinity variance.
Paper Structure (16 sections, 8 equations, 3 figures, 3 tables)

This paper contains 16 sections, 8 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Approach Summary
  3. Related Work
  4. Machine Learning for Spectral Analysis in Digital Soil Mapping
  5. Multimodal Fusion for Soil Salinity Estimation
  6. Comparative Analysis of Cross-Domain Adaptation Methods
  7. Methodology
  8. Teacher Model Pretraining [RQ-1]
  9. Spectral Domain Adaptation - [RQ2, RQ3]
  10. Multi-modal Knowledge Distillation [RQ-2]
  11. Performance Benchmarks & Discussion
  12. Dataset and Area of Interest
  13. Data Preprocessing
  14. Experimental Setup
  15. Model Performance Evaluation & Ablation Studies
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Comprehensive architecture of the DeepSalt framework. (\ref{['fig:overview']}) illustrates the end-to-end workflow: pretraining a teacher on lab FTIR data, spectral adaptation, and distilling knowledge into a student that fuses adapted satellite spectra with ancillary environmental data. (\ref{['fig:model-123']}) details the Spectral Adaptation Unit (SAU), which bridges the domain gap between laboratory (FTIR, MIR) and satellite (EnMAP, VNIR-SWIR) spectra by projecting them into a shared latent space. (\ref{['fig:training']}) shows the multi-objective distillation strategy that transfers the teacher's feature representations and output distributions to the student model.
  • Figure 2:
  • Figure 3: