Sources of Error in Synthetic Remote Sensing Data and Potential Impacts on Ecohydrological Models in Semiarid Rangelands

Abstract:

Semiarid rangelands have a high level of both spatial and temporal vegetation heterogeneity due to slow net primary production rates and highly variable rainfall. Ecohydrological modeling in these ecosystems requires high resolution inputs of vegetation structure and function. We used the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) to create eight synthetic Landsat TM images across a growing season (April - September). STARFM fuses the high spatial resolution of Landsat TM with the high temporal resolution of Terra MODIS. Previous attempts to assess the accuracy and quantify model errors of STARFM have used pixel-based regression and difference image analysis, as well as examining the distribution of those errors across land cover types. However, those model errors have not previously been compared to a null model (i.e., using the nearest available Landsat scene). If there is very little change occurring, then you would expect the model to have artificially high correlation coefficients and low error estimates. Additionally, we examined several other potential sources of error: i) time of year or season, ii) vegetation height class from airborne LiDAR, iii) solar radiation (i.e., aspect), and iv) snow. We found that STARFM added new information when compared to the null model, yet the null model was highly accurate during large parts of the growing season (June through September, r² = 0.95 - 0.97) suggesting that simply reporting r² values from pixel-based regression is insufficient to assess model accuracy. We found that areas with snow in the preceding model input imagery (NDSI > 0.4) increased errors threefold (RMSE(snow) = 0.3223, RMSE(not-snow) = 0.1017). We also found that pixels with shrub or tree vegetation (height > 0.3 m) tended to have higher errors when compared to ground or grass pixels. Finally, our results indicate that during the middle of the growing season, there are patterns in the error that relate to solar radiation with the NDVI of tree and grass pixels being underestimated on south-facing slopes and overestimated on north-facing slopes. Shrub pixels show a more consistent overestimation of NDVI.