Hyperspectral remote sensing and inversion of forestry biophysical parameters

Hyperspectral Applications

The absorption characteristics in the near-infrared region of the spectrum (1000–2500 nm) are a function of the bending and stretching vibrations of biochemical bonds (e.g., between hydrogen and carbon atoms and nitrogen and oxygen atoms) and their harmonics and overtones. In the visible region, chlorophyll and carotenoids produce strong absorption due to electronic energy transitions. Different methods exist for retrieving leaf and canopy characteristics from reflectance measurements.

(1) Empirical models (exponential and/or multiple regression):The principle involves constructing mathematical combinations of multiple reflectance measurements across narrow or broad spectral bands and correlating them with specific features of the observed surface. These relationships are calibrated based on experimental or simulated reflectance databases (built upon radiative transfer models). This type of approach is straightforward but has some limitations: when calibrated to an experimental database, the representativeness of the relationships is limited to the database's representativeness (Figure 1). Furthermore, exponential and multivariate regressions can be sensitive to multiple individual features. They are also sensitive to atmospheric conditions, imaging geometry, and spatial resolution, thus typically requiring calibration for each image.

△Figure 1. Strategies for extracting forest attributes from remote sensing data within the framework of forest research

(2) Model inversion:This method uses a model to simulate the reflectance spectra of leaf, canopy, and soil characteristics, and then obtains leaf and canopy parameters through model inversion. Typical numerical inversion methods involve fixing some parameters to reasonable values ​​and determining other parameters (free parameters) by minimizing a function of the difference between observed reflectance (Pmeas) and simulated reflectance (Pmod). This function is called the evaluation function. An efficient minimization method, such as the simplex optimization method proposed by Nelder and Mead (1965), was recommended by Privette et al. (1994) after testing several different inversion methods. For practical applications, inversion techniques based on pre-computed reflectance databases are generally preferred over computationally more expensive iterative methods. Frequently used efficient computational methods include lookup tables and neural networks.

Both of the above methods rely on training databases based on simulations or measurements. While uncertainties in measurements and models can lead to significant variations in results, numerical inversion strategies must also address other challenges: the inversion problem involves determining the causes of expected or observed results, or calibrating mathematical model parameters to reproduce the observations. Inversion problems typically do not satisfy the important assumptions of well-posedness: there may be no strictly defined solution, the solution may not be unique, and/or it may be non-consistently dependent on the data. Inversions of such models often yield a large number of different possible solutions, i.e., they are ill-conditioned. Therefore, using effective regularization strategies (constraining the minimization of the empirical error function) is crucial for numerical inversion strategies.

△Figure 2. Collection of hyperspectral forestry biophysical parameters