Spatial Object Definition for Vegetation Parameter Estimation from Hymap Data

Common per-pixel estimations for vegetation parameters are hampered by spatial mismatch between the image and ground observations, and limited by neglecting spatial patterns. Geometric correction of images can reach accuracies in the range of 1 pixel, while locations of ground observations are measured with an accuracy of 5-10m by GPS. Our HyMap image has 5m pixels. Consequently, although coordinates may match, ground observations are not necessarily linked to the correct pixel, but can undesirably be represented by neighbouring pixels. Furthermore, vegetation patterns define the observation units used by ecologists, but they are not reflected by square pixels. Even though these patterns may reveal useful information, it is excluded from the analysis. Object-oriented image analysis offers significant improvements. Objects are formed by groups of spectrally similar, neighbouring pixels; this reduces the risk of spatial mismatch. They are thus believed to provide a better approach to vegetation-parameter estimation than the conventional per-pixel approach. Objects are defined by spectral similarity, but an important question is how much spectral variance is allowed. The aim of this paper is to investigate optimal heterogeneity for predicting biomass and LAI. We have data from 250 field plots in our test site, 60 km west of Montpellier in southern France. A HyMap image is available as well. The image is segmented with different heterogeneities; larger heterogeneities resulting in larger segments. Field observations are linked to corresponding objects and with Ridge regression, relations between field observations and reflection values are identified. For each heterogeneity the prediction error is determined; the smallest error indicating optimal heterogeneity. Conclusions confirm that increasing the object size shows an optimum in prediction accuracy for both biomass and LAI. * Corresponding author.