Spatiotemporal reconstruction of MODIS land surface temperature with the help of GLDAS product using kernel-based nonparametric data assimilation

Abstract. Spatiotemporal inconsistency in MODIS land surface temperature (LST), one of the most widely used geospatial data parameters, is a concern for its application in various studies. Moreover, there are limited methods that can address spatiotemporal reconstruction of LST in diverse physiography. The use of kernel-based spatiotemporal assimilation in a multitemporal approach to reconstruct LST in a complex physiographic region, northeast India, is addressed. Global land skin temperature (Ts) from Global Land Data Assimilation System (GLDAS) is chosen as a temporal covariate for MODIS LST due to its consistent temporal resolution. Considering the high temporal correlation between GLDAS Ts and MODIS LST, temporal assimilation is done in the first stage followed by spatial assimilation in the second stage. Due to data gaps, a kernel-based nonparametric estimator is adopted to map the spatiotemporal distribution and reconstruct LST with spatiotemporal consistency. This approach shows satisfactory performance in restoration of spatial variation in the study area with a coefficient of determination (R2  =  0.98), root mean square error (RMSE = 0.61 K), and absolute value of bias (B  =  0.22  K). Also, the credibility of reconstructed LST from this method is found to be as good as the baseline data (MODIS LST) given that there is no severe data deficiency with uneven spatial availability in the baseline data itself. Comparison of gap-filled LST showed positive correlation with ground-based measurements (average R  =  0.71), reflecting decent agreement with seasonality. Hence, the kernel-based nonparametric assimilation could be used to reconstruct LST using a multitemporal approach in complex physiographic regions, which could be used to fill the spatiotemporal data gaps and increase data consistency in MODIS LST.