Daily estimates of Landsat fractional snow cover driven by MODIS and dynamic time-warping

Detecting changes in understorey and canopy vegetation cycles in West Central Alberta using a fusion of Landsat and MODIS

“…MODIS EVI data values were spatially smoothed using a 3 × 3 low‐pass filter to eliminate spatial noise. Results were calculated with 50% weighting on the centre pixel and 50% on the neighbourhood (Berman et al, ). Annual vegetation curves were then obtained through cubic spline interpolation (Melaas et al, ) and these were further smoothed using a Gaussian filter with a sigma value of 4.…”

“…Theoretically, the warp path possibilities are exponentially high; however, a rule set was generated based on the path which minimized the Euclidean distance between target and query MODIS EVI, with curves starting at day of the year 1 and ending at day 365. The algorithm used a 50‐day window in which to restrict the resultant rule sets such that if observations were >50 days apart, a rule set was not able to form a relation (Berman et al, ). This second step was repeated until all years had become a target year generating a rule set for every year.…”

“…Target year observations were not rearranged as they signify real values. In situations where multiple values were warped to the same day of year, the mean of all values for that day was used (Berman et al, ). Landsat EVI vegetation curves were created for every year by iteratively generating rulesets with that year becoming the target year and having all other years (query years) warped to said target year.…”

“…In order to model inter-annual change at moderate spatial scales more accurately, one alternative is to fuse Landsat and MODIS data sets. Dynamic Time Warping (DTW) is a fusion technique that was first developed as a speech recognition algorithm (Sakoe & Chiba, 1978) and has been proven capable of determining patterns in corresponding remote-sensing data sets (Baumann et al, 2017;Berman et al, 2018;Petitjean, Inglada, & Gançarski, 2012). It has the ability to increase temporal density of historic time series by shifting annual dates based on rule sets generated from complimentary datasets.…”

“…They determined that, compared to ground camera data, the DTW Landsat product was more accurate than MODIS alone at predicting transition dates. Berman et al () adapted these techniques to accurately assess inter‐annual change in regional snow cover at a fine spatial scale with greater than 80% accuracy when compared to ground cameras. Results were further used to model grizzly bear response to varying snow conditions (Berman, Coops, Kearney, & Stenhouse, ).…”