Using linear spectral unmixing of high spatial resolution and hyperspectral data for geometric-optical modelling

Abstract

The linear spectral mixing model is a widely used technique in remote sensing to estimate the fractions of several individual surface components present in an image pixel and the pure reflectance spectrum of a component, called endmember, is the model’s necessary parameter. Different methods can be used to extract endmembers, finding out the pure pixel from hyperspectral data is the most common way, but at the regional scale, the selected single pixel can not present the typical component accurately. The objective of this paper is to estimate the feasibility of up-scaling from high spatial resolution data to medium spatial resolution hyperspectral data by linear spectral unmixing based technique for extracting the required endmembers. In this case, an inverted Li-Strahler geometric-optical model is applied to retrieve one of the forest canopy variables (crown closure) in a broadleaved forest natural reserve, located in the Three Gorges region of China. This model needs three important scene components (sunlit canopy, sunlit background and shadow). The three components’ classes are firstly estimated using QuickBird fusional image with 0.6 m spatial resolution by eCognition, an object-based classification method. Then, a 50 by 50 pixels moving window calculated each component’s proportion is matched to individual pixels with 30 m spatial resolution of EO-1 Hyperion data. The umixing model is finally used for deriving the three endmembers based on the Hyperion data with surface reflectance and the per-pixel three fractions computing from QuickBird classification. From the spectral profile comparison, the scaling-based endmembers indicate the mean spectra of the components unlike the pure pixels’ spectra. For validating the results, we use 32 independent sample sites collected in the study area to assess the accuracy of inverted model’s outputs, and the results of scaling-based endmember extraction method (R²=0.60) seem better than the pixel-based method (R²=0.51).