By Martin van Leeuwen.
Abstract
The effect of viewing geometry on the retrieval of the fractional cover of maize (Zea Mays L.) has been investigated by applying linear spectral mixture analysis on AHS imagery (in the spectral range of 400 to 1042 nm) acquired at view zenith angles ranging from 15° to 46°. Imagery of the crop fields was acquired for forward and backscattering of sunlight. The maize fields in the imagery exhibited profound bidirectional reflectance effects, hence complicating the retrieval of the fractional cover. Linear unmixing was achieved using Monte Carlo simulation, with n = 100, in order to propagate endmember variability. Two sets of Monte Carlo linear unmixing were compared: (1) Monte Carlo Unmixing (MCU) using nadir vegetation endmembers that were simulated using the PROSAIL canopy reflectance model, and (2) a novel approach that made use of vegetation endmembers that were simulated under the respective view zenith angles of the pixels to be unmixed (AMCU). The results showed that differences in vegetation cover estimates exceeded 10% cover (α = 0.05) for 24.5% of the pixels in the image exhibiting backscattering of sunlight, with a noticeable increase in difference between estimates for larger viewing zenith angles. In the forward scatter direction, none of the pixels showed a difference in cover estimate exceeding 10% cover. The AMCU method showed better consistency in fractional cover estimates for geographically corresponding pixels from the backscatter and forward scatter image. Also, lower RMSE values were found for the AMCU method when its results were validated against ground truth cover measurements.
Keywords: linear unmixing, Monte Carlo unmixing, spectral mixture analysis, viewing geometry, view zenith angle, sun-target-sensor geometry, PROSAIL, bidirectional reflectance, BRDF