Fig 1.
Flowchart for super-resolution process.
The star symbols are where our refinement from the original algorithm [26] was implemented. As an example, the super-resolution process for converting MODIS 500-m resolution images (band 3–7) to 250-m resolution images is shown.
Fig 2.
Original super-resolution algorithm proposed by Tonooka in 2005 [26].
Fig 3.
Proposed super-resolution algorithm.
Fig 4.
Reference satellite data and land cover map for the study site.
(Left) False color image taken by ASTER (2001/09/24), (center) that taken by MODIS, and (right) JAXA land cover map degraded to 250-m resolution. All images have a UTM 54 projection with a WGS84 ellipsoid. Land cover category abbreviations: DBF, deciduous broadleaf forest; DNF, deciduous needleleaf forest; EBF, evergreen broadleaf forest; ENF, evergreen needleleaf forest.
Table 1.
Characteristics of MODIS bands and correspondence with reference ASTER bands.
Fig 5.
Simulated Point Spread Function (PSF) for 100×100 pixels (i.e., ν = 100).
(left) Triangular weighting function for sensor PSF, (center) Gaussian weighting function for optical PSF, and (right) combined PSF for a Moderate Resolution Imaging Spectroradiometer (MODIS) observation.
Fig 6.
Tuning of the regularization parameter λ in Algorithm 3.
(left column) Root mean squared error (RMSE) and correlation coefficient (CC) for a wide range (from 0 to 104) and (right column) RMSE and CC for a narrow range (from 10−4 to 2.0×10−2). Both tunings were performed with the first super-resolution image (i.e., retrieval of 500-m thermal images), and the common λ was used for the second super-resolution process. Each x-axis is log-scale, whereas each y-axis is scaled by the RMSE or CC for λ = 0. The dashed line marks λ = 0.002.
Table 2.
Correlation Coefficient (CC), Root Mean Squared Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR) for each band [31, 32], and Relative Dimensionless Global Error (ERGAS) between the results from the three types of super-resolution algorithms and ASTER radiance.
CC and PSNR: larger is better; RMSE and ERGAS: smaller is better.
Table 3.
Basic statistics over the entire study region between the results from the three super-resolution algorithms, original MODIS image (1-km resolution), and ASTER image.
Fig 7.
Improvement in spatial resolution of MODIS thermal infrared bands by the proposed super-resolution Algorithm 3.
(left column) Original 1000-m resolution images, (center column) result from the first super-resolution process, and (right column) result from the second super-resolution process for (top row) MODIS band 31 and (bottom row) MODIS band 32. The radiance value was converted into brightness temperature by Eq 10.
Fig 8.
Comparison of maps obtained by the three algorithms.
The upper panel containing 8 images displays MODIS band 31 results, and the lower one displays MODIS band 32 results. For each panel, from the far left column: reference ASTER images, MODIS retrieved by Algorithm 2, Algorithm 3, and Algorithm 1, respectively. The top row shows the 500-m resolution retrieval (first super-resolution images for Algorithms 2 and 3), and the bottom row shows the 250-m resolution retrieval. Algorithm 1 (original algorithm) cannot retrieve a 500-m resolution map. The red pixels (encircled by red circles for visibility) in the maps retrieved by Algorithm 1 indicate negative brightness temperatures.
Table 4.
Statistics for super-resolved thermal data and accuracy indices for Algorithm 3 for each land-cover type.
The land-cover type was determined using previously reported data [34], while unclassified and snow/ice categories were excluded. Note that the four forest categories (see Fig 4) were integrated into the one category. Radiance (W/m2/str/μm) with standard deviation, Tb (brightness temperature; K) with standard deviation, PSNR, ERGAS, and N (number of sample pixels) are listed.
Table 5.
Comparison of accuracy indices (ERGAS: Smaller is better; PSNR: Greater is better) for different algorithms and land-cover types.
Fig 9.
Maps describing characteristics of super-resolution retrieval.
From the top row, pixel homogeneity, Mahalanobis distance, and data sources (the typical spectral pattern or neighboring pixel used in the retrieval process) are displayed. The left and center columns show the maps for the first and the second super-resolution retrievals by our proposed Algorithm 3, respectively. The right column shows retrieval by the original Algorithm 1.