Spatial upscaling of remotely sensed leaf area index based on discrete wavelet transform

Leaf area index (LAI), a crucial parameter of vegetation structure, provides key information for Earth's surface process simulations and climate change research from local to global scale. However, when the LAI retrieval model built at local scale (high resolution) is directly applied at a large scale (low resolution), a spatial scaling bias may be caused. The magnitude of this bias depends on the non-linearity of retrieval model and heterogeneity of land surface. Various spatial upscaling algorithms have been developed to correct for this scaling bias. In this study, we try to explore the potential application of wavelet transform in spatial upscaling. Hence, an algorithm based on the relation between the bias rate in scaling and the detail lost rate in discrete wavelet transform (DWT) was proposed to eliminate scaling bias at a large scale. To evaluate the proposed algorithm, three sites with different degrees of heterogeneity from Validation of Land European Remote Sensing Instruments database were chosen. Using Systeme Probatoire d'Observation dela Tarre, Operational Land Imager, and corresponding ground measurements, the performances of the proposed algorithm were further quantitatively analysed. Additionally, the upscaling accuracy between the algorithm based on Taylor Series Expansion (TSE) and that based on DWT was compared. Generally speaking, the root mean square error (RMSE) and relative error (RE) of retrieved LAI induced by the scale bias can be greatly reduced after correction with those two algorithms. Over high heterogeneous landscape, the upscaling performance is more obvious. When the corresponding synchronous priori knowledge is available, the proposed DWT-based algorithm has reached a comparative accuracy with the TSE-based algorithm. The RE can decrease from 13.54% to 3.47% and RMSE from 0.36 to 0.09 over the selected heterogeneous landscape. When the synchronous priori knowledge is not available, the proposed DWT-based algorithm outperforms the TSE-based algorithm. The RE and RMSE can decrease from 22.98% and 0.49 to 7.97% and 0.13, respectively. However, unlike the TSE-based algorithm, the proposed DWT-based algorithm is simpler and not constrained by the characteristic of the retrieval model. These results indicate that it is feasible to successfully correct for the scaling bias by using the proposed DWT-based spatial upscaling algorithm.