A neural network inversion of the DART model to retrieve Norway spruce LAI at a very high spatial resolution

Abstract

Leaf Area Index (LAI) is a key input parameter in many eco-physiological and climate models. Therefore, the development of methods to accurately and timely retrieve LAI over large areas is essential to fully understand the Earth system. The inversion of radiative transfer models is a “universal” method to retrieve LAI from remotely sensed images because it is independent from the study area and the sampling conditions. In this paper, we study the potential of the 3D Discrete Anisotropic Radiative Transfer (DART) model to retrieve the LAI of a Norway spruce forest stand. An extensive airborne/field campaign was carried out in September 2004 to acquire AISA Eagle VNIR hyperspectral images (pixel size of 0.4 m) and to collect ground truth data for the image pre-processing, DART parameterization and validation of the LAI estimations. Because DART is a complex and computationally demanding model, it was first run in direct mode to build a large dataset of possible canopy realizations. Then, a relatively simple two-layer feed forward backpropragation neural network was trained using the simulated DART top of canopy reflectances and a priori information on canopy closure. Finally, the LAI inversion was performed over the radiometrically and atmospherically corrected AISA Eagle images by using a sliding window whose size matches the extent of the DART modelled forest scenes. Results indicate that the inversion of the DART model to retrieve the LAI of complex Norway spruce canopies using ANN is a promising tool. Nevertheless, the approach still has to be improved in case of very high spatial resolution images.