Bionote
Doctoral researcher in high-throughput plant phenotyping, currently exploring the combination of 3D radiative transfer modelling and deep learning to estimate biochemical traits from crop canopies at leaf scale. My research focuses on bridging simulations of 3D virtual scenes with drone-based optical imagery to advance scalable and accurate vegetation monitoring.
My background includes a BSc in Agronomic Engineering and an MSc in Geospatial Technologies, complemented with six years of applied industry experience implementing remote sensing solutions and advanced spatial analysis for precision agriculture applications.
Presentation Abstract
This study introduces a drone-based High-Throughput Phenotyping (HTP) approach that combines 3D radiative transfer modeling and deep learning to estimate Leaf Chlorophyll Content (LCC) from simulated multispectral imagery of different crops. HTP has emerged as a faster, more efficient, and scalable alternative to traditional field phenotyping, yet it poses challenges in data generation, processing, and equipment constraints, often resulting in limited annotated datasets. While some studies have employed radiative transfer modeling for phenotyping, many fail to fully account for the 3D structural heterogeneity of complex canopies. To address this, we use the 3D Discrete Anisotropic Radiative Transfer (DART) model to generate realistic virtual scenes of maize (Zea mays L.) and sugar beet (Beta vulgaris L.) canopies, incorporating structural complexity and spatial variability through parameterized optical properties for plants and soil. From these 3D scenes, we simulate high-resolution (1 cm) multispectral drone imagery across 10 spectral bands. These synthetic images—labeled with known leaf-level biochemical traits—are used to train a esNet-like U-Net model for per-pixel LCC estimation. The model is then tested on unseen simulated datasets with varying canopy structures, growth stages, and scene conditions. Our results demonstrate strong performance, achieving up to R² = 0.83 and RMSE = 6.5 µg/cm² in pixel-level LCC estimation, with robust ccuracy across variations in LCC distribution, canopy structure, and row orientation. Current efforts focus on extending the method to additional growth stages and conditions to evaluate its transferability and robustness against structural, biochemical, and environmental shifts.
Although not yet validated on real UAV imagery, these in silico results highlight the method’s potential for estimating chlorophyll content in individual leaves within complex, heterogeneous crop canopies.
Keywords: 3D radiative transfer, machine learning, high-throughput phenotyping, UAV, multispectral data, leaf biochemical traits.
Co-Authors: Zbynek Malenovsky, Ribana Roscher