A Spectral Angle Distance-Weighting Reconstruction ...

0 downloads 0 Views 1MB Size Report
Mar 21, 2013 - Abstract—Land surface temperature (LST) is an important pa- rameter in the physics ... Emissivity product MOD11A2 with 46 time series data in total ..... [7] Z. X. Tan, S. Liu, B. K. Wylie, C. B. Jenkerson, J. Oeding, J. Rover, and.
1514

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 9, SEPTEMBER 2014

A Spectral Angle Distance-Weighting Reconstruction Method for Filled Pixels of the MODIS Land Surface Temperature Product Tong Shuai, Xia Zhang, Shudong Wang, Lifu Zhang, Member, IEEE, Kun Shang, Xiaoping Chen, and Jinnian Wang

Abstract—Land surface temperature (LST) is an important parameter in the physics of land surface processes, but a large number of pixels are often filled as zero due to cloud, heavy aerosols, and so on in the Moderate-resolution Imaging Spectroradiometer (MODIS) LST product. This letter presents the spectral angle distance (SAD)-weighting reconstruction (SADWER) method of reconstruction of zero-filled pixels of the MODIS LST product. It relies on the hypothesis that pixels with the same land-cover type have nearly the same LST in a localized area. SAD can measure the similarity of land-cover types of different pixels, and pixels with higher land-cover similarity can contribute more to the reconstruction using the weighting method. The result shows that the reconstruction ratio could be as high as 95% using only the SADWER method and nearly 100% after spatial filter postprocessing. The reconstruction accuracy is validated using artificially generated 20-, 50-, and 80-km-diameter concentrically filled areas in both forest and crop land-cover types. The statistical result shows that the standard deviations of the reconstruction errors are less than 2 Kelvin.

LST is affected by latitude, solar altitude angle, elevation, wind, and so forth, at the global scale; however, LST is directly related to the land surface cover type in a localized region [23]. For example, vegetation or water can decrease the surface temperature, whereas urban or barren areas can increase it. Additionally, LST is related to the proportion of different land-cover types. A significant relationship between LST and fractional vegetation cover (FVC) is indicated by the regressive function of LST against FVC [24]. Spectral angle distance (SAD) is a key indicator that measures the spectral similarity of two pixels in the remote sensing image [25]. Thus, SAD can be used to measure land-cover similarity. Based on the idea of reconstructing the zero-filled LST pixels using pixels with high land-cover similarity in a localized region, a SAD-weighting reconstruction (SADWER) method for the MODIS LST Product is proposed in this letter.

Index Terms—Filled pixels, land surface temperature (LST), moderate-resolution imaging spectroradiometer (MODIS), reconstruction, spectral angle distance (SAD).

II. R ECONSTRUCTION M ETHOD

I. I NTRODUCTION

L

AND surface temperature (LST) is one of the key parameters in the physics of land surface processes as well as in global warming and climate change monitoring on local through global scales [1]. LST data have been widely applied in many research fields, such as studies of plateaus [2], cities [3], soil [4], pre- or postfire [5], [6] or wildfire assessments [7], hydrology [8], evapotranspiration [9], drought [10], arctic areas [11], lithological mapping [12], phenology [13], air condition [14], permafrost [15], earthquake [16], and others. The Moderate-resolution Imaging Spectroradiometer (MODIS) LST product is one of the most widely used LST products [17]–[19]. However, a large number of pixels of this product are filled as zero due to cloud, heavy aerosols, and so on in the imaging scene [20]. Until now, few studies have focused on reconstructing the filled pixels of the MODIS LST product [21], [22]. Manuscript received October 4, 2013; revised December 9, 2013; accepted December 27, 2013. This work was supported by the Global Change Comparative Study of Planets and Earth (Y3SG1900CX) and the National Natural Science Foundation of China (41371359). The authors are with the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100101, China (e-mail: zhangxia@radi. ac.cn). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2013.2297735

A. Experimental Area Haihe River Basin in North China was chosen as the experimental area in this study. There are various land-cover types in this basin, such as croplands, forests, shrublands, grasslands, wetlands, urban and built-up areas, and so forth (Fig. 1). The terrain of this basin, which has an area of approximately 319 000 km2 , mainly comprises plains, mountains, and plateaus from low elevation in the southeast to high elevation in the northwest. B. Data Three MODIS products throughout the year of 2005 are used in this research: the MODIS 8-day Land Surface Temperature/ Emissivity product MOD11A2 with 46 time series data in total (LST daytime data, LST daytime quality control data); the MODIS 8-day Surface Reflectance (SR) product MOD09A1 (surface reflectance bands 1–7 data, state flags data); and the MODIS Land Cover Type product MCD12Q1 (land-cover type 1-IGBP classification). As the spatial resolution of LST Product is 1 km, the other two products with 500-m spatial resolution are spatially resampled into the same spatial resolution with LST product. The SR data are used to identify pixels that have the same or extremely similar land-cover type compared with the filled pixel. The LST quality control data are used to ensure the LST quality of the pixels used for reconstruction, whereas the SR state flags data are used to ensure that the cloud condition of those pixels is clear.

1545-598X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

SHUAI et al.: SADWER METHOD FOR FILLED PIXELS OF MODIS LST PRODUCT

1515

Fig. 2. LST reconstruction flow. “LST Mandatory QA Flags = 00 or 01” means that LST is produced. “LST Data Quality Flag = 00” means good data quality for the L1B data in seven TIR bands. “MOD35 cloud  00” means that the cloud state of SR is not clear.

Fig. 1. Land-cover types (MODIS MCD12Q1, 2005) and location of Haihe River Basin.

C. Reconstruction Flow Our reconstruction method is proposed based on the following hypothesis: pixels with the same land-cover type have nearly the same LST in a localized area. This hypothesis has been substantiated by previous papers [23]. For each LST filled pixel (FP), we first use the MODIS Land Cover Type product to identify preliminary reference pixels (RPs) having the same land-cover type as the FP. Then, the reflectance SADs of the FP and RPs are calculated to filter the RPs, and these SADs are then used for reconstruction as weighting parameters so that pixels with higher land-cover similarity contribute more toward the reconstruction. After WER, a spatial filtering postprocessing step is implemented to reconstruct isolated filled pixels. The detailed processing steps (Fig. 2) are described as follows: Step 1) Identification of FPs in the remote sensing image. According to the description of MODIS LST data, the valid range of the MODIS LST product is 7500–65 535 (scale factor is 0.02). Thus, pixels with LST lower than 7500 are considered FPs. Step 2) Identification of preliminary RPs for each FP that meet the following three requirements: 1) The RPs have the same land-cover type as the FP. 2) The distance between RP and FP is relatively short (100 km is set specifically in this study).

Fig. 3. Comparison of results before and after SADWER. A is band 10 (14–21 March) and B is band 25 (12–19 July). For consistency with the MODIS LST product, the scale factor is set to 0.02, and the LST unit is Kelvin (the same below).

1516

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 9, SEPTEMBER 2014

Fig. 4. Reconstruction results of band 25 after spatial filtering 1, 3, and 5 times (A, B, and C, respectively).

3) The LST quality of RPs is good: (A) LST is produced (LST mandatory QA flags = 00 or 01); (B) the quality of the L1B data in seven thermal infrared (TIR) bands used for retrieving the MODIS LST product is rated as “good” (LST data quality flag = 00); and (C) the LST error of RPs is ≤ 2 Kelvin. Step 3) SADWER: reconstruct LST with reflectance SADs of multitemporal data as weighting parameters. The pixels of the MODIS LST product are filled mostly due to clouds or other atmospheric conditions at the time (the ith 8-day time series); thus, the SR of FP and RP between 8-day time series numbers (i − 5) and (i + 5) is extracted first for calculating SADs. Then, cloud-affected SR time-series data of the 11 8-day time series are discarded according to the reflectance state data (MOD35 cloud = 00). After that, the SAD [25] of FP and RP is calculated (1) using the remaining reflectance data on cloudclear days. If the SAD is lower than the similarity threshold (ST), 0.1 in this study, the RP is included in the weighting group. If the weighting group is empty after all of the RPs identified in Step 2 are processed according to the above instructions, nothing is done to the FP. Otherwise, WER of the FP’s LST is implemented according to (2) and (3) ⎛ ⎞ nb  ⎜ ⎟ F Pi RPi ⎜ ⎟ i=1 −1 ⎜ ⎟ (1) SAD = cos ⎜  1  1 ⎟   2 2 nb nb ⎝  ⎠  F Pi2 RPi2 i=1

i=1

where nb = number of cloud-clear bands, and i = band i LSTF P =

N RP s

(ωj × LSTj )

(2)

j=1

ST − SADj ωj = N RP s (ST − SADj )

(3)

j=1

where LSTF P = the reconstructed LST of the FP, NRP s = the number of RPs in the weighting group, LSTj = the LST of the jth RP , ωj = the weighting parameter for LSTj , ST = the similarity threshold,

Fig. 5. Total number of filled pixels and the reconstruction ratio of 46 bands after WER and five reiterations of spatial filtering.

and SADj = the SAD of FP and the jth RP in the weighting group. Step 4) Spatially filter isolated or small-area FPs. After SADWER, isolated or small-area FPs may remain. Assume that the LST of one point and that of the adjacent points are highly correlated. As the spatial variation in LST is relatively slight, the adjacent pixels can be used to reconstruct the FP using the spatial filtering method. In this letter, the eight adjacent pixels are defined as the up, down, left, right, up-left, up-right, down-left, and down-right pixels. If the number of valid LST adjacent pixels is equal to or greater than three, these valid LSTs will be averaged as the reconstructed LST for this FP. To reconstruct small-area FPs, Step 4 can be implemented several times. However, because the reconstruction error will increase as the processing times of spatial filtering increase, the number of reiterations is set to five in this study.

III. R ECONSTRUCTION R ESULTS A. Results of SADWER Method After WER, most filled pixels belonging to popular land-cover types are reconstructed (Figs. 3 and 5). The reconstruction ratios of bands 10 and 25 in Fig. 3 are around 95% and 90%, respectively. Furthermore, it is worth mentioning that the reconstructed LST pixels are texturally continuous with the original LST pixels. No obvious signs of artificial reconstruction between reconstructed and the original LST pixels are apparent.

B. Results After Spatial Filtering After the SADWER process, some small-areal or isolated FPs still exist [Fig. 3(b)]. Those FPs are well reconstructed using the spatial filtering method (Fig. 4). As shown in Fig. 5, the reconstruction ratio of each band is higher than 70% after the first spatial filter, and in most cases, it increases to almost 100% after the fifth spatial filter (e.g., bands 5–14, 23–29, 37–41, etc.). The reconstruction ratio of the first spatial filter is higher than WER in some bands, such as bands 15–16, 19–22, 30, 32, because the filled pixels in those bands are relatively few in number (fewer than 400), and most are isolated pixels that could be reconstructed using the spatial filtering method.

SHUAI et al.: SADWER METHOD FOR FILLED PIXELS OF MODIS LST PRODUCT

1517

Fig. 8. Reconstruction ratio and the statistics of the reconstruction errors (difference between the original LST values of the MODIS product and the reconstructed LST values) of the northern (N) and the southern (S) artificially filled areas. The number after “N” or “S” indicates the diameter of the filled area. TABLE I M EAN AND S TANDARD D EVIATION OF R ECONSTRUCTION E RRORS FOR M AJOR L AND -C OVER T YPES (U NIT: K ELVIN )

Fig. 6. Location and reconstruction results of artificially (A) 20-, (B) 50-, and (C) 80-km-diameter concentrically filled areas (band 20). Numbers 1, 2, 3, and 4 represent the artificially filled image, the SADWER result image, the first spatial filtering image, and the original LST product image, respectively. N indicates the northern artificially filled area, and S indicates the southern area.

Fig. 7. Spatial distribution of the reconstruction errors after the first filter. A, B, and C represent the 20-, 50-, and 80-km artificially filled areas, respectively. N and S indicate the northern and the southern areas, respectively. The scale factor of LST reconstruction errors is converted to 1 for easy understanding, and the unit is still Kelvin (the same below).

IV. R ECONSTRUCTION VALIDATION The previous section shows that more than 90%, and as many as 100%, of the filled pixels can be reconstructed by successively applying the SADWER and spatial filtering methods. The reconstruction accuracy is validated below. To validate the reconstruction accuracy at different spatial scales and for different land-cover types, we artificially generated 20-, 50-, and 80-km-diameter concentrically filled areas in both forest (northern) and crop (southern) land-cover areas of the Haihe River Basin (Fig. 6). The original band 20 of the LST product was chosen as the experimental data because this band has very few filled pixels, as can be observed in Fig. 6. Line No. 2 in Fig. 6 demonstrates the WER result of the artificially filled areas. It shows that the overwhelming major-

ity of the filled pixels are reconstructed. Compared with the original LST product (Line No. 4), the reconstructed textures of the filled areas are very clear and exactly correspond to the originals. In particular, for the 80-km-diameter filled areas, this strongly shows that the SADWER method based on land-cover similarity is adequate for LST reconstruction of relatively large area filled pixels. As most of the filled pixels were reconstructed by the SADWER method, only a few isolated filled pixels are left to be spatially filtered. From Line No. 3 in Fig. 6, we see that after the first spatial filtering, all of the isolated filled pixels are reconstructed. Spatial filtering acts as an efficient postprocessing method after WER. The spatial distribution of the LST reconstruction errors of artificially filled areas is illustrated in Fig. 7. The reconstruction ratio and error statistics for different areas are plotted in Fig. 8, and the reconstruction errors for major land-cover types are demonstrated in Table I. As the artificially filled area grows to both the north and the south, the ratio of WER decreases and reconstruction error increases. Specifically, the mean and standard deviation of the reconstruction errors increase slightly or remain almost the same, whereas the extrema (maxima/minima) rise/drop dramatically. However, most of the mean ± standard deviation values keep within ±2 Kelvin, indicating that the reconstruction errors of the artificially filled areas are generally lower than 2 Kelvin. For most land-cover types, the mean values of reconstruction errors are lower than 0.5 Kelvin, and the standard deviations are less than 2. Moreover, compared with the southern reconstruction results, the reconstruction ratio of the northern area is lower, and the reconstruction error is larger. This may be caused by the complexity of the land cover and the land terrain in the

1518

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 9, SEPTEMBER 2014

forest-type region, which is much greater than those in the croptype region. V. C ONCLUSION AND D ISCUSSION From the above reconstruction results and the validation with artificially filled data, we conclude that the SADWER method proposed in this letter is adequate for reconstructing relatively large filled pixels for the MODIS LST products. It can, moreover, be applied in similar LST products, such as those retrieved from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Advanced Very High Resolution Radiometer (AVHRR), and others. As a postprocessing and complementary method, spatial filtering is a necessary step for LST reconstruction because many isolated pixels cannot be reconstructed with the SADWER method. However, the spatial filtering method is very effective with this kind of filled pixels. The MODIS Land Cover Type product is used in this study to help increase the efficiency of searching pixels with the same land cover. However, the Land Cover Type product is not necessary. It can help to narrow the search, but the SAD determines the reference LST pixels. Thus, if no Land Cover Type product can be used, the SAD can be used as the criterion for searching pixels with the same land cover, although this will require more processing time. Along with its strengths, some caveats in the application of the SADWER method should be considered. 1) In this letter, we set the reference range to 100 km, and pixels within this range are considered relevant to the filled pixel. The reference range is set according to the land surface conditions of the experiment area. If the land surface is relatively simple and smooth, the reference range can be set larger; otherwise, it should be set smaller. 2) If the filled area is much larger than the reference range, the SADWER method cannot reconstruct the filled LST well. 3) The effect of elevation on LST should be considered if the terrain dramatically changes. R EFERENCES [1] Z. L. Li, B.-H. Tang, H. Wu, H. Ren, G. Yan, Z. Wan, I. F. Trigo, and J. A. Sobrino, “Satellite-derived land surface temperature: Current status and perspectives,” Remote Sens. Environ., vol. 131, pp. 14–37, Apr. 15, 2013. [2] M. S. Salama, R. Van der Velde, L. Zhong, Y. Ma, M. Ofwono, and Z. Su, “Decadal variations of land surface temperature anomalies observed over the tibetan plateau by the Special Sensor Microwave Imager (SSM/I) from 1987 to 2008,” Clim. Change, vol. 114, no. 3/4, pp. 769– 781, Oct. 2012. [3] B. Huang, J. Wang, H. Song, D. Fu, and K. K. Wong, “Generating high spatiotemporal resolution land surface temperature for urban heat island monitoring,” IEEE Geosci. Remote Sens. Lett., vol. 10, no. 5, pp. 1011– 1015, Sep. 2013. [4] X. N. Song, P. Leng, X. T. Li, X. H. Li, and J. W. Ma, “Retrieval of daily evolution of soil moisture from satellite-derived land surface temperature and net surface shortwave radiation,” Int. J. Remote Sens., vol. 34, no. 9/10, pp. 3289–3298, May 1, 2013. [5] S. Veraverbeke, W. W. Verstraeten, S. Lhermitte, R. Van De Kerchove, and R. Goossens, “Assessment of post-fire changes in land surface temperature and surface albedo, their relation with fire-burn severity using multitemporal MODIS imagery,” Int. J. Wildland Fire, vol. 21, no. 3, pp. 243–256, 2012. [6] G. Guangmeng and Z. Mei, “Using MODIS land surface temperature to evaluate forest fire risk of Northeast China,” IEEE Geosci. Remote Sens. Lett., vol. 1, no. 2, pp. 98–100, Apr. 2004.

[7] Z. X. Tan, S. Liu, B. K. Wylie, C. B. Jenkerson, J. Oeding, J. Rover, and C. Young, “MODIS-informed greenness responses to daytime land surface temperature fluctuations and wildfire disturbances in the Alaskan Yukon River Basin,” Int. J. Remote Sens., vol. 34, no. 6, pp. 2187–2199, Mar. 2013. [8] F. Silvestro, S. Gabellani, F. Delogu, R. Delogu, and G. Boni, “Exploiting remote sensing land surface temperature in distributed hydrological modelling: The example of the Continuum model,” Hydrol. Earth Syst. Sci., vol. 17, no. 1, pp. 39–62, Jan. 2013. [9] S. M. Bateni, D. Entekhabi, and F. Castelli, “Mapping evaporation and estimation of surface control of evaporation using remotely sensed land surface temperature from a constellation of satellites,” Water Resourc. Res., vol. 49, no. 2, pp. 950–968, Feb. 2013. [10] N. T. Son, C. F. Chen, C. R. Chen, L. Y. Chang, and V. Q. Minh, “Monitoring agricultural drought in the Lower Mekong Basin using MODIS NDVI and land surface temperature data,” Int. J. Appl. Earth Observ. Geoinf., vol. 18, pp. 417–427, Aug. 2012. [11] A. Soliman, C. Duguay, W. Saunders, and S. Hachem, “Pan-arctic land surface temperature from MODIS and AATSR: Product development and intercomparison,” Remote Sens., vol. 4, no. 12, pp. 3833–3856, Dec. 2012. [12] T. J. Majumdar, S. K. Pal, and A. K. Bhattacharya, “Generation of emissivity and land surface temperature maps using MODIS TIR data for lithological mapping over the Singhbhum-Orissa Craton,” J. Geol. Soc. India, vol. 80, no. 5, pp. 685–699, Nov. 2012. [13] J. M. Hanes and M. D. Schwartz, “Modeling land surface phenology in a mixed temperate forest using MODIS measurements of leaf area index and land surface temperature,” Theor. Appl. Climatol., vol. 105, no. 1/2, pp. 37–50, Aug. 2011. [14] W. B. Zhu, A. L˜u, and S. Jia, “Estimation of daily maximum and minimum air temperature using MODIS land surface temperature products,” Remote Sens. Environ., vol. 130, pp. 62–73, Mar. 15, 2013. [15] S. Hachem, M. Allard, and C. Duguay, “Using the MODIS land surface temperature product for mapping permafrost: An application to Northern Quebec and Labrador, Canada,” Permafrost Periglacial Process., vol. 20, no. 4, pp. 407–416, Oct.–Dec. 2009. [16] S. K. Panda, S. Choudhury, A. K. Saraf, and J. D. Das, “MODIS land surface temperature data detects thermal anomaly preceding 8 October 2005 Kashmir earthquake,” Int. J. Remote Sens., vol. 28, no. 20, pp. 4587– 4596, Oct. 2007. [17] Z. M. Wan, “New refinements and validation of the MODIS land-surface temperature/emissivity products,” Remote Sens. Environ., vol. 112, no. 1, pp. 59–74, Jan. 15, 2008. [18] R. Zorer, D. Rocchini, M. Metz, L. Delucchi, F. Zottele, F. Meggio, and M. Neteler, “Daily MODIS land surface temperature data for the analysis of the heat requirements of grapevine varieties,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 4, pp. 2128–2135, Apr. 2013. [19] C. Y. Wu and Z. Niu, “Modelling light use efficiency using vegetation index and land surface temperature from MODIS in Harvard Forest,” Int. J. Remote Sens., vol. 33, no. 7, pp. 2261–2276, Apr. 2012. [20] S. H. Shen and G. G. Leptoukh, “Estimation of surface air temperature over Central and Eastern Eurasia from MODIS land surface temperature,” Environ. Res. Lett., vol. 6, no. 4, p. 045206, Oct.–Dec. 2011. [21] L. H. Ke, C. Q. Song, and X. Ding, “Reconstructing complete modis lst based on temperature gradients in Northeastern Qinghai-Tibet Plateau,” in Proc. IEEE IGARSS, 2012, pp. 3505–3508. [22] M. Neteler, “Estimating daily land surface temperatures in mountainous environments by reconstructed MODIS LST data,” Remote Sens., vol. 2, no. 1, pp. 333–351, Jan. 2010. [23] P. Hesslerova, J. Pakorný, J. Brom, and A. Rejšková-Procházková, “Daily dynamics of radiation surface temperature of different land cover types in a temperate cultural landscape: Consequences for the local climate,” Ecol. Eng., vol. 54, pp. 145–154, May 2013. [24] Y. S. Zhang, O. A. Inakwu, and E. Ramadan, “Assessment of land surface temperature in relation to landscape metrics and fractional vegetation cover in an urban/peri-urban region using Landsat data,” Int. J. Remote Sens., vol. 34, no. 1, pp. 168–189, Jan. 2013. [25] F. A. Kruse, A. B. Lefkoff, J. W. Boardman, K. B. Heidebrecht, A. T. Heidebrecht, P. J. Barloon, and A. F. H. Goetz, “The Spectral Image-Processing System (Sips)-interactive visualization and analysis of imaging spectrometer data,” Remote Sens. Environ., vol. 44, no. 2/3, pp. 145–163, May/Jun. 1993.