Object-based retrieval of biophysical canopy variables using artificial neural nets and radiative transfer models

Abstract

The retrieval of biophysical variables using canopy reflectance models is hampered by the fact that the inverse problem is ill-posed. This leads to unstable and often inaccurate inversion results. In order to regularize the model inversion, a novel approach has been developed and tested on synthetic Landsat TM reflectance data. The method takes into account the neighbouring radiometric information of the pixel of interest, named object signature. The neighbourhood data can either be extracted from gliding windows, already segmented images, or using digitized field boundaries. The extracted radiometric data of the neighbourhood pixels are used to calculate 42 descriptive statistical properties that comprehensively characterize the spectral (co)variance of the image object (e.g. mean and standard deviation of the distributions, intercorrelations between spectral bands, etc.). Together with the habitual spectral signature of the pixel being inverted (6 variables), this object signature (42 variables) is used as input in an artificial neural net to estimate simultaneously three important biophysical variables (i.e. leaf area index, leaf chlorophyll, and leaf water content). The use of neural nets for the model inversion avoids time-consuming iterative optimizations and provides a computational effective way to consider simultaneously pixel and object signatures. In order to “learn” the relation between spectral signatures and biophysical variables, the neural nets were previously trained on large synthetic data sets. The data sets consist of pixel signatures and the corresponding signatures of image objects representing various agricultural fields. The signatures were simulated with the SAILH+PROSPECT canopy reflectance model, assuming largely varying intra- and interfield distributions of the model input parameters. To demonstrate the benefits of the object-based inversion, neural nets were also trained on the pixel signatures alone. For this purpose, the object signatures were simply replaced by randomly generated white noise—all other conditions being the same. The intercomparison based on 30,000 independent validation patterns showed that the proposed method significantly enhances the estimation accuracies: for example, the leaf area index (LAI) is estimated with a percental root mean square error (PRMSE) of 18.3% (object-based) compared to 25.1% (pixel based); the corresponding numbers for the leaf chlorophyll content are 12.6% compared to 15.9%; for the equivalent leaf water thickness, 10.6% and 13.9%, respectively. The benefit of the object signature was strongest for the LAI. Concerning this important biophysical variable, the novel concept accounted for almost one-half of the remaining unexplained variance of the traditional pixel-based approach. Increased accuracies were attributed to the fact that intrafield variations of biophysical canopy variables lead to object signatures that are modulated by the actual average leaf angle (ALA) of the canopy. Since ALA can be considered constant insight given an agricultural field, the concurrent use of pixel and object signatures significantly reduces confounding effects between LAI and ALA typical for traditional inversion approaches. Compared to competing approaches, the algorithm can also be applied to monotemporal imagery and does not require a priori information or the identification of crop type.