Self-Supervised Deep Learning Methods for Hyperspectral Image Analysis of Soil Properties in Agricultural Applications

Abstract:

Imaging spectroscopy (hyperspectral imaging) provides continuous spectral measurements ideal for estimating crop traits and soil properties at scale; however, its agricultural application is limited by a lack of labelled data, the high dimensionality of hyperspectral data, and domain shift between sensors and regions. This thesis investigates self-supervised learning and hybrid regression architectures for hyperspectral analysis of soil properties in agricultural applications.

The first contribution, HyperKon, is a contrastive learning framework that includes a hyperspectral-native convolutional backbone, spectral squeeze-and-excitation blocks, hard negative mining, and a hyperspectral perceptual loss. Pretrained on EnHyperSet-1 (curated from 800 spaceborne EnMAP scenes with 224 bands from 420-2450 nm) using spectral-consistent augmentations, HyperKon outperformed baselines on pansharpening and classification benchmarks, yielding representations that transfer with minimal fine-tuning.

Building on these learned representations, our second contribution HyperSoiNet, combines a HyperKon encoder with an ensemble of Random Forest, XGBoost, and k-nearest neighbours for soil property estimation. On the HyperView airborne field spectroscopy challenge dataset (150 bands, 462-938 nm), using stratified cross-validation and temporally separated splits, HyperSoilNet achieves coefficients of determination up to R² = 0.786 for phosphorus and R² = 0.771 for potassium, with lower performance for pH (R² = 0.529), and reduces error by 23.8% relative to baselines.

To further exploit spectral structure, our third contribution SpecBPP, introduces spectral band permutation prediction, a spectrum-aware objective where a network recovers the ordering of permuted spectral segments under a gradually harder curriculum. For SOC regression, SpecBPP pretraining achieves R² = 0.9456 and residual prediction deviation (RPD) = 4.19, outperforming masked autoencoding and joint-embedding baselines, particularly in low-label regimes.

These results show that incorporating spectral physics into self-supervised objectives and hybrid predictors results in more accurate, data-efficient models capable of scalable soil monitoring for sustainable agriculture.