Conference Agenda

Microscopic Hyperspectral Imaging (MHSI) is vital for analyzing the link between morphology and physiology in rice pathology. However, hardware limitations often result in low spatial resolution, blurring fine structures like stomata. This issue hinders the accurate assessment of physiological stress, especially when distinguishing rice genotypes with similar visual symptoms. To address this, we propose PhytoSR, a super-resolution network with a decoupled spatial-spectral architecture. Specifically, we design a Bidirectional 3D Spectral Mamba Branch. This branch uses a bidirectional scanning mechanism to capture global spectral dependencies efficiently (with linear complexity), ensuring spectral accuracy. In parallel, a 2D Texture Branch is used to recover fine spatial details. Extensive experiments on a self-constructed rice dataset show that PhytoSR significantly outperforms SOTA methods. Comprehensive ablation studies further validate the necessity and effectiveness of the bidirectional Mamba mechanism and the decoupled design. Furthermore, we developed a Stomatal Disease Resistance Index (SDRI) based on the reconstructed images. This index enables the non-invasive quantification of subcellular physiological damage, revealing distinct immune responses among susceptible rice varieties.

Imaging Spectroscopy, or multispectral/hyperspectral imaging, combines image processing and spectroscopy. Despite proven effectiveness in agriculture, its practical adoption remains limited.

A major constraint hindering widespread adoption in agriculture is that spectral cameras remain expensive, bulky, and difficult to scale for routine use. Despite emerging lower-cost systems, a gap remains in cameras that meet the practical demands of commercial crop monitoring.

A key gap is the lack of explicit 3D shape information to account for geometric effects, as plant reflectance varies with viewing and lighting.

Another research gap is that analysis methods currently treat spatial and spectral information largely independently, which risks losing important interactions between plant morphology and reflectance patterns.

A barrier to advancing spectral imaging with deep learning is the lack of large, annotated, and diverse datasets.

Finally, despite interest in converting RGB to spectral representations, information theory shows true reconstruction of spectral signatures from RGB is infeasible, especially for physiological and biochemical signals.

This presentation will review the current state of the art and focus on emerging solutions, with concrete examples of initiatives such as the EU HyperImage project. We will conclude with an outlook on future directions needed to translate these advances into robust and scalable practice.

An accurate assessment of plant nutritional status is essential for the sustainable production of citrus crops. This study advances non-destructive diagnostic methods by combining Vis/NIR hyperspectral imaging with machine learning techniques to predict macro- and micronutrient concentrations in spring flush leaves of Clementina de Nules mandarin collected from commercial orchards throughout two growing seasons. A total of 576 hyperspectral images of these leaves were acquired under laboratory conditions within a spectral range of 400–980 nm. Two hyperspectral-based approaches were evaluated: (i) analysis of mean reflectance spectra using multiple machine learning algorithms, and (ii) application of convolutional neural networks (CNNs) to full hyperspectral images. Using the mean-spectral approach, the Multilayer Perceptron (MLP) outperformed the other algorithms, achieving coefficients of determination (R²) above 0.60 for most nutrients, with maximum values of 0.85 for Fe. CNN-based models showed comparable performance and, in some cases, outperformed traditional machine learning approaches when trained on subsets of wavelengths. Hyperspectral models were also compared with conventional vegetation indices, which exhibited poor predictive performance. Overall, these results highlight the potential of hyperspectral imaging combined with machine learning as a robust and scalable tool for real-time nutrient monitoring, supporting optimised fertilisation management in citrus production through non-destructive means.

Spectral reconstruction represents a promising strategy to obtain hyperspectral information from low-cost imaging sensors. This study evaluated the reconstruction of vineyard hyperspectral data acquired under real field conditions through a deep learning-based spectral super-resolution approach. Reference hyperspectral cubes in the VisNIR range were collected using a proximal hyperspectral camera operating outdoors. After radiometric calibration and reflectance normalization, the data were pre-processed and divided into training, validation, and independent test sets.

Spectral reconstruction was implemented using a deep residual convolutional neural network (HSCNN-R), designed with stacked 2D convolutional layers and residual skip connections to preserve spatial consistency and enhance convergence stability.

The reconstructed spectra showed very high agreement with the reference measurements, with coefficients of determination close to 0.98 and mean reconstruction errors in the order of only a few percentage points across the VisNIR range. Key absorption features related to plant biochemical components were consistently preserved, and the spectral shape was accurately reproduced over the full wavelength interval.

These results confirm the technical feasibility of spectral super-resolution in vineyard environments and support the integration of low-cost RGB sensors into precision viticulture systems, significantly reducing instrumentation complexity and operational costs while maintaining high spectral reliability.

Low-temperature stress poses a significant threat to the growth and quality of tea, yet traditional physiological measurements often lack the timeliness and efficiency needed for effective stress assessment. To address this challenge, we propose an intelligent cold stress monitoring framework that integrates hyperspectral imaging (HSI) with long short-term memory (LSTM) time-series modeling. Cold stress experiments were conducted on tea plants under conditions of 0°C, -5°C, and -10°C. HSI along with physiological measurements of moisture, relative conductivity (RC), and chlorophyll, were collected simultaneously. A novel physiological composite index, the Cold Damage Index (CDI), was developed to quantitatively assess cold stress severity. Compared to traditional methods (PLS and SVM), the LSTM-based models demonstrated significantly higher prediction accuracy. By incorporating multi-head band attention and time attention mechanisms, the proposed multi-head band + time attention (MBTA)-LSTM model achieved R² values of 0.9821 for RC and 0.9778 for moisture, with chlorophyll predictions showing R² values ranging from 0.91 to 0.95. Using CDI as the classification target, the model achieved 97.88% accuracy in categorizing cold damage into three levels of severity. This study demonstrates the feasibility of combining hyperspectral time-series predictions with a unified physiological index for the early detection of cold-induced stress.

Withering is a critical process that significantly influences the flavor, color, and overall quality of white tea. Traditional evaluation of withering degree mainly relies on manual sensory judgment, which is subjective and inefficient for rapid and accurate monitoring. To address this limitation, this study proposes an evaluation method based on hyperspectral imaging and machine learning to assess withering degree and quality changes during the withering process. By integrating spectral and image features extracted from hyperspectral images, a classification model was developed to identify the withering degree of white tea, achieving a classification accuracy of 93.33%. Moisture content, tea polyphenols, chlorophyll, and catechins were selected as representative quality indicators, and quantitative models were developed for rapid and non-destructive prediction of major quality components in white tea, with R² greater than 0.80 for all indicators. In addition, hyperspectral imaging was applied to visualize the spatial distribution of key quality indicators, enabling dynamic characterization of chemical changes during the withering process. The results demonstrate that the proposed method can effectively achieve rapid and non-destructive evaluation of the white tea withering process, providing theoretical support and technical guidance for online monitoring and refined quality control in white tea processing.