Balancing prediction accuracy and generalization ability: A hybrid framework for modelling annual dynamics of satellite-derived land surface temperatures
Abstract
Satellite-derived land surface temperatures (LSTs) are vital in land surface processes. However, the presence of cloud and some other image-specific problems generates data gaps at various aperiodic times of the year, which severely constrain the analysis of remotely sensed time series. Many annual temperature cycle (ATC) models had been developed to produce spatio-temporally seamless LSTs. However, a flexible ATC framework that simultaneously considers the prediction and generalization ability and that provides options for users under various practical needs remains lacking. Considering these limitations, we proposed a flexible ATC framework, based on which we developed a hybrid ATC (ATCH) model by balancing model complexity and generalization ability. The ATCH was modeled by adding a sinusoidal function (ASF) and several meteorological and surface factors including surface air temperature (SAT), normalized differential vegetation index (NDVI), albedo (ALB), relative humidity (RHU) and soil moisture (SM) on the traditional single sinusoidal ATC model (named ATCO). ATCH model was tested on MODIS LST data over mainland China, and evaluations were conducted with two recently popular gapfilling methods (i.e., RKI and RSDAST) and in-situ data. The results show that the overall RMSE of the ATCH model decreases about 1.77 K and 0.71 K over mainland China in day and night when compared with the ATCO, respectively. The accuracy improvement varies with the land cover types with greater improvements over the more vegetation areas. The performance of ATCH model also exhibits significant seasonal variations, with the RMSEs during summer and winter higher than those during the other two seasons. Further evaluations reveal that the ASF, SAT, NDVI and other meteorological and surface factors provide the contribution around 16%, 40%, 15% and 30% on accuracy improvement, respectively. ATCH exhibits a better performance when compared with other gapfilling methods. Quantitative evaluation results indicate that ATCH model can reconstruct LST under cloud with mean RMSE of 3.18 and 2.16 K in day and nighttime, respectively. Though with limitations, we consider our model can make up for several key limitations in existing ATC models, improve the quality of LST products and further support the associated applications.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFMGC31J1367L
- Keywords:
-
- 1616 Climate variability;
- GLOBAL CHANGEDE: 1632 Land cover change;
- GLOBAL CHANGEDE: 1637 Regional climate change;
- GLOBAL CHANGEDE: 1640 Remote sensing;
- GLOBAL CHANGE