Daily snow depth data – 61,480 pairs of observed values provided by 145 ground stations – were selected as samples for data quality evaluation. The data span 424 time phases in two snow seasons over Tibetan Plateau from October 1, 2009 to April 30, 2011. The confusion matrix and formulas (1) and (2) were adopted to evaluate the snow classification images and analyze the MODIS data (MOD10AI, MYD10A1) under clear weather. The accuracy of all types of synthetic products during cloud processing, including general classification accuracy and snow classification accuracy, was analyzed. For this purpose, the following samples were considered during accuracy evaluation: ① sample (a) with snow images and records from ground stations (snow depth > 0); ② sample (b), recorded snow from ground stations classified as snow-free, namely omission; ③ sample (c), recorded snow-free from ground stations classified as snow; and ④ sample (d) with snow-free images and records from ground stations (Table 3).
Confusion matrix for accuracy verification
General classification accuracy, namely accuracy, reflects the capacity of classification algorithms in identifying snow as snow-covered and land as snow- free across the whole research area. Snow classification accuracy, namely precision, reflects the proportion of real snow pixels among all the snow pixels identified by the classification algorithm. They may be expressed by the following formulas:
Raw data of MOD10A1, MYD10A1 and all processed products in the confusion matrix are indicated in Figure 4. The general classification accuracy of MODIS data was above 98%, while the snow classification accuracy approximated to 82%. Since there were few pixel elements modifying the erroneous judgment of lake and lake ice, statistical results were omitted. The following factors might result in errors during precision validation: thin snow cover, dispersed snow distribution, differentiated spatial scales employed for generating snow depth data at ground stations and MODIS snow image pixels, optics’ penetrating effect in thin snow-covered areas, and spectral mixing effect caused by mottled thin snow. When snow cover is thin (for example, less than 3 cm) during snow-melt periods, snow underestimation might occur, in which circumstance snow is observed within the small area of ground stations but the MODIS image (500 m * 500 m) displays to be snowless. It might also be that the station is located in a city and the snow quickly melts due to the temperature which is higher than its surrounding areas, in which circumstance the MODIS image overestimates the situation and erroneously suggests snow. In the case where snow is thicker than 3 cm, the samples recorded as snow by both images and ground observations are denoted by a3, and the samples recorded as snow by images but as land by ground observations are denoted by b3. A calculation by formulas (1) and (2) demonstrated an improved general accuracy of all synthetic products, with Oa and Sa of MODIS data respectively rising to about 99% and 94%.
Table 4 shows that the cloud removal procedure prior to elevation filtering had less influence on the classification accuracy; daily observation combination and elevation filtering slightly improved the general snow classification precision, whereas maximum snow and land masking brought down the general classification accuracy of synthesis products; in the case of the data with snow- depth being above 3 cm, the accuracy decreased to 98.24% from 98.97% (average accuracy value of MOD10A1 and MYD10A1) while the snow classification accuracy increased from 94.48% to 94.73%, with a larger amount of cloud removed; fitting expected snow lines had the lowest accuracy and was suitable for the last step of cloud removal; nevertheless, when the snow depth was above 3 cm, the general classification accuracy and snow classification accuracy of cloud-free snow products still reached 96.6% and 89% respectively, higher than previously published documentary records24 and basically met the accuracy requirements of MODIS standard snow products under clear weather. All of this indicates that the algorithm for cloud removal may be applied in the dynamic, daily, cloud-free snow monitoring over Tibetan Plateau.
Accuracy verification of all product types
|Accuracy of general|
|Accuracy of snow|
All sta- tions
Snow dep- th ≤ 3 cm removed
All sta- tions
Snowdep- th ≤ 3 cm removed
|Adjacent three- day temporal synthesis|
|“Permanent” snow and land recognition|
|Neighboring four -pixel method|
|7||Elevation fil- tering||4,059||722||495||42,761||3,656||151||97.47||98.63||84.90||96.03|
|Maximum snow and land masking|
|Expected snow lines fitting (cloud- free s now products)|