Reflectance correction in tree shadows in high spatial resolution imaging spectroscopy using radiative transfer simulations and machine learning
Correction de la réflectance dans les ombres des arbres en imagerie a haute résolution spatiale spectroscopie utilisant des simulations de transfert radiatif et l'apprentissage automatique
Résumé
With the development of high spatial resolution imaging spectroscopy, getting an accurate surface reflectance retrieval is a crucial issue in deriving a quantitative assessment of relevant physical variables used in the fields of land cover mapping, soil, vegetation and cultures monitoring or characterization of impervious surfaces condition in urban areas for instance. One hindering factor is the treatment of shadows. Indeed, 3D atmospheric correction models already successfully manage opaque shadows by accounting for the buildings and the terrain. But, they mostly fail for shadows caused by semi transparent medias such as tree ones. As part of the irradiance in the shadow is directly transmitted through the tree crown and comes from the multiple scaterrings with it, there is a need to properly account for it in the radiative budget modelled in atmospheric correction codes. This transmitted irradiance depends on many parameters that could be divided in 3 parts : (1) Scene parameters : tree properties (e.g. dimensions, leaf optical properties - LOPs), environment (e.g. soil and aerosol) and illumination conditions (e.g. sun zenithal angle), (2) Spatial parameters : position of the pixel in the shadow relatively to the tree location (3) spectral parameter : spectral bands. A past study designed a reflectance correction in tree shadows by working on two scene parameters (1) namely the sun zenithal angle and the leaf area index (LAI). But it was based on a reference tree model with fixed dimensions and only 3 different LOPs. Then concerning the spatial (2) and spectral (3) parameterization, a mean spectral tree crown transmittance was predicted and weigthed by a tree viewing angle depending on the pixel position to finally aim for building a correction factor. From these aforementioned limitations, a new study is ongoing to avoid the previous separability hypothesis between spectral and spatial dependencies in the computation of the correction factor, by experimenting new regression techniques from machine learning approaches and taking into account both tree structural and leaf biochemical traits (instead of LOPs). Therefore, based on (1), (2) and (3) variability, there is a need to have the best comprehension of the complex radiation phenomenon happening in tree crowns and their shadows. We first carried out 3D physical simulations with the radiative transfer code DART, in order to perform a sensibility analysis, based on Sobol’s indices computation, highlighting key scene parameters. We have built different metamodels based on gaussian processes and polynomial regressions considering both spectral and spatial variation to predict the tree crown transmittance. The treatment of spatial and spectral information was made using Karhunen-Loeve decomposition or autoencoders, reducing the data dimension and the computation time simultaneously. Hence, it’s not necessary to train a specific metamodel for each spectral band and pixel but only a few ones to retrieve global dependencies. Finally, we obtain a new spatial and spectral dependent correction factor wich will be tested on airborne hyperspectral images. The results will be compared to those obtained with the past tree shadow correction model.