Combining MODIS and AMSR-E based vegetation moisture retrievals for improved fire risk monitoring

Research has shown that remote sensing in both the optical and microwave domain has the capability of estimating vegetation water content (VWC). Though lower in spatial resolution than MODIS optical bands, AMSR-E microwave measurements are typically less affected by clouds, water vapor, aerosol or solar illumination, making them complementary to MODIS real time measurements over regions of clouds and haze. In this study we explored a wavelet based approach for combining vegetation water content observations derived from higher spatial resolution MODIS and lower spatial resolution AMSR-E microwave measurements. Regression analysis between AMSR-E VWC and spatially aggregated MODIS NDII (Normalized Difference Infrared Index) was first used to scale MODIS NDII to MODIS VWC products. Our approach for combining information from the two sensors resorts to multiresolution wavelet decomposition of MODIS VWC into a set of detail images and a single approximation image at AMSR-E resolution. The substitution method of image fusion is then undertaken, in which the approximation image is replaced by AMSR-E VWC image, prior to using inverse wavelet transform to construct a merged VWC product. The merged VWC product thus has information from both MODIS and AMSR-E measurements. The technique is applied over low vegetation regions in Texas grasslands to obtain merged VWC products at intermediate resolutions of ~1.5km. Apart from offering a way to calibrate MODIS VWC content products to AMSR-E observations, the technique has the potential for downscaling AMSR-E VWC to higher spatial resolution over moderately cloudy or hazy regions where MODIS reflective bands become contaminated by the atmosphere. During such situations when contaminated MODIS signals cannot be used to obtain the wavelet detail images, MODIS detail images from a preceding time step is used to downscale the current AMSR-E VWC to higher resolutions. This approach of using detail images from the recent past would be justified if the detail images containing the high frequency components of the image change slowly. Correlation analysis of detail images from consecutive time steps shows that this is approximately true, at-least for the low spatial resolution detail images. Our approach yields accuracy of around 77% on the average over the selected study region and temporal period. This technique thus has the potential for ensuring the data continuity of high spatial resolution VWC products, a requirement essential for fire risk monitoring.