Maize Leaf Biomass Retrieval at Multi-growing Stage Using UAV Multispectral Images Based on 3D Radiative Transfer Process-guided Machine Learning

High-precision estimation of above-ground biomass (AGB) is very crucial for breeding field. Optical variables (e.g. vegetation index (VI)) have widely used in monitoring AGB. In this study, we used a stem-leaf separation strategy to estimate biomass. The combined use of multispectral and deep learning techniques (e.g., convolutional neural network (CNN) and transfer learning (TL)) estimate leaf biomass (LGB). Then, an allometric growth model was used to estimate stem biomass (SGB). We used three-dimensional radiative transfer (3D RTM) - LESS model to simulate a universality and realistic multispectral dataset $(n=44880)$. We designed a CNN architecture that can extract multi-layer feature of CNNs. This study combined 3D RTM dataset, CNN and TL techniques to estimate maize LGB of multiple growth stage. The results showed our method had the best performance in LGB estimation at multiple growth stage. The result showed that using the allometric growth model to estimate SGB achieved an $R^{2}$ of 0.83 and an RMSE of $67.5 \mathrm{~g} / \mathrm{m}^{2}$, improving the prediction accuracy of SGB. This study utilized the advantage of 3D RTM, CNN, TL, and allometric model which can monitor maize AGB more accurately for breeding filed.