Adaptive Clustering for Efficient Phenotype Segmentation of UAV Hyperspectral Data

TL;DR

Hyperspectral data from UAVs are rich but voluminous, challenging real-time processing on edge devices. The authors propose OHSLIC, an online clustering framework coupled with a lightweight 1D-CNN classifier to perform per-line tree phenotype segmentation and predict leaf contents such as $\text{chlorophyll}$, $\text{carotenoid}$, and $\text{anthocyanin}$ from synthetic 224-band spectra generated with PROSPECT-D. A high-fidelity simulator-based dataset (Blender + PROSPECT-D) and an adaptive clustering mechanism enable real-time edge processing, with the OHSLIC-C-C variant using confidence-driven refinement to improve segmentation and regression. Experimental results on a Jetson Nano show that OHSLIC-C-C achieves competitive Dice and superior $R^2$ for the three-parameter regression while maintaining low per-line inference times (~$12.4$ ms) compared with pixel-based baselines, demonstrating the method’s practical utility for scalable vegetation monitoring and drought/forest-health applications. Overall, the work advances edge-enabled hyperspectral analysis by integrating adaptive online clustering with compact neural networks and realistic synthetic data for robust on-device phenotyping.

Abstract

Unmanned Aerial Vehicles (UAVs) combined with Hyperspectral imaging (HSI) offer potential for environmental and agricultural applications by capturing detailed spectral information that enables the prediction of invisible features like biochemical leaf properties. However, the data-intensive nature of HSI poses challenges for remote devices, which have limited computational resources and storage. This paper introduces an Online Hyperspectral Simple Linear Iterative Clustering algorithm (OHSLIC) framework for real-time tree phenotype segmentation. OHSLIC reduces inherent noise and computational demands through adaptive incremental clustering and a lightweight neural network, which phenotypes trees using leaf contents such as chlorophyll, carotenoids, and anthocyanins. A hyperspectral dataset is created using a custom simulator that incorporates realistic leaf parameters, and light interactions. Results demonstrate that OHSLIC achieves superior regression accuracy and segmentation performance compared to pixel- or window-based methods while significantly reducing inference time. The method`s adaptive clustering enables dynamic trade-offs between computational efficiency and accuracy, paving the way for scalable edge-device deployment in HSI applications.

Figures (10)

• Figure 1: Illustration of the proposed real-time processing technique for the data of a hyperspectral push-broom camera for on-board phenotype segmentation of trees.
• Figure 2: Distributions of the PROSPECT-D model parameters for the leaves in the generated data. This is to give an idea of the distribution of the labels in the dataset and of what values the different parameters can take. These are the values for chlorophyll, carotenoid, brown content, water thickness, dry matter, and anthocyanin respectively.
• Figure 3: Example simulation data: fake RGB image on the top left, with spectra inside the highlighted squares on the top right. Best viewed in colour. The bottom to panel show the pixel labels for the trees. These contain values of 0 or 1 for if the tree is of certain class or not.
• Figure 4: Illustrative figure of the model and the pipeline. OHSLIC takes a line as input and gives a set of clusters to the classifier network with multiple heads for the multiple predictions. The classifier network then gives its confidence back to OHSLIC and also outputs the labels for the different clusters. The number of clusters used in OHSLIC can be controlled by the clusters controller.
• Figure 5: Illustrative figure of the Classifier model. The model is made of a 1D CNN feature extractor and 4 heads, which are MLPs of different sizes. The output of the model is a 1D array of length 3 for the three different parameters and the confidence of the segmenation.