Neural network method to solve inverse problems for canopy radiative transfer models
- PDF / 554,473 Bytes
- 12 Pages / 595.276 x 793.701 pts Page_size
- 45 Downloads / 201 Views
NEURAL NETWORK METHOD TO SOLVE INVERSE PROBLEMS FOR CANOPY RADIATIVE TRANSFER MODELS1
UDC 681.32
A. N. Kravchenko
Vegetation parameter retrieval is considered as the inverse of modeling canopy radiative transfer. To solve this problem, a new computationally efficient method based on mixture density networks (MDNs) is proposed to estimate the errors of retrieved parameters for each given set of reflectances. The properties of neural networks of traditional architecture and MDNs are considered. The method is tested using a simple model and the PROSPECT leaf radiative transfer model and is validated against real data. Keywords: mixture density network, canopy radiative transfer model, inverse problems, Earth remote sensing, plant moisture. INTRODUCTION Data on soil and plant moisture are used for vegetation monitoring and early drought prediction, which is especially important under fast climatic changes. Advanced devices for Earth remote sensing (ERS) and ERS data available have made it possible to determine the state of vegetation in global and regional scales [1]. At a local level, applying remote methods provides a basis for noninvasive characterization of plants [2]. Determining plant moisture is a part of a more general problem: using satellite data to retrieve architectural and biochemical characteristics of crop canopy, such as leaf area index, fraction of absorbed radiation, content of pigments, primarily chlorophyll. A modern and promising trend is the assimilation of satellite measurements into dynamic prediction models (for example, crop growth models to solve crop prediction and control problems [3]) and heat- and moisture-transfer model in the soil–plant–atmosphere system to monitor droughts, predict high water and weather, and to model the climate [4, 5]. Modern data assimilation techniques [6, 7] impose two significant requirements on ERS data processing methods: measurement error should be estimated and the method should be computationally efficient. The latter requirement is due to the necessity of processing large data flows (billions pixels per day for modern satellite systems). The methods available for estimating canopy parameters do not meet the above-mentioned requirements. In particular, computationally efficient methods based on vegetative indices and artificial neural networks with traditional architecture do not provide information on the error in parameter estimates. Variational and Monte Carlo methods are computationally inefficient, the computational efficiency of lookup table methods substantially depends on the table size and the number of input parameters. Therefore, creating a new computationally efficient method for finding canopy parameters (in particular, moisture content) that provide information on the error of parameter estimates is topical and necessary for modern canopy growth models and models of heat and moisture transfer in the soil–plant–atmosphere system. 1
The paper is supported by the joint project of the Science and Technology Center in Ukraine (STCU) and National
Data Loading...
