cuSTARFM is a GPU-enabled Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM)
Version 1.0
MODIS and Landsat surface reflectance products have complementary characteristics in terms of spatial and temporal resolutions. To fully exploit these datasets, the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) was developed by Gao et al. (2006). The STARFM approach blends the high-frequency temporal information from MODIS and the high-resolution spatial information from Landsat to generate synthetic surface reflectance products at 30m spatial resolution and daily temporal resolution. STARFM uses one or more pairs of Landsat-MODIS images collected on the same dates to predict the surface reflectance at Landsat resolution on other MODIS observation dates. However, the computational performance of STARFM has been a bottleneck for mass production, even with the parallel option in STARFM v1.2 (Geo, et al., 2015; Gao, et al., 2017).
To overcome the computational barrier and support mass production of large-size images, we designed and implemented a GPU-enabled STARFM program based on the Compute Unified Device Architecture (CUDA), called cuSTARFM. By taking advantages of the large amount of concurrent computing threads of a GPU, cuSTARFM can greatly reduce the computing time and improve the computational performance. Experiments showed that cuSTARFM achieved a speedup of 342 using a Nvidia Tesla K40 GPU, compared with a sequential STARFM program running on an Intel Xeon E3-1226 CPU.
STARFM_PARAMETER_START
#The number of input pairs of Landsat-MODIS images (>=1)
NUM_IN_PAIRS = 1#The pf band of Landsat
The_pf_band_of_Landsat_for_calculating = 3#The pc band of MODIS
The_pc_band_of_MODIS_for_calculating = 1#The input MODIS images
#When NUM_IN_PAIRS > 1, multiple images must be given
#File names are separated by space
IN_PAIR_MODIS_FNAME = D:\data\newdata\MODO9A1.A2009249.dat#The input Landsat images
#When NUM_IN_PAIRS > 1, multiple images must be given
#File names are separated by space
IN_PAIR_LANDSAT_FNAME = D:\data\newdata\2009-249-flaash.dat#The MODIS images for the prediction dates
#Multiple images can be given
#File names are separated by space
IN_PDAY_MODIS_FNAME = D:\data\newdata\MODO9A1.A2009329-0.dat D:\data\newdata\MODO9A1.A2009329-0.dat#The output synthetic prediction images
#Multiple images can be given
#File names are separated by space
OUT_PDAY_LANDSAT_FNAME = D:\data\newdata\2009-329new-flaash-test2.tif D:\data\newdata\2009-329new-flaash-test3.tif#The_width of searching_window
The_width_of_searching_window = 31#Assumed_number of classifications
Assumed_number_of_classifications = 4#The relative importance of space distance
The_relative_importance_of_space_distance = 25#Landsat sensor error
Landsat_sensor_error = 20#MODIS sensor error
MODIS_sensor_error = 50#Output image format (optional)
#Will be used when the extension of the output files
#is not given
G_Type = GTIffSTARFM_PARAMETER_END
Note: MODIS and Landsat images use different band number sequence. You can relate them using the following table:
Landsat(pf) | MODIS(pc) |
---|---|
1 | 3 |
2 | 4 |
3 | 1 |
4 | 2 |
5 | 6 |
7 | 7 |
The program runs as a command line. You may use the Command (i.e., cmd) in Windows, or a terminal in Linux/Unix.
Note: The computational performance of cuSTARFM largely depends on the GPU. The more powerful is the GPU, the better performance.