CMVs are determined by comparing consecutive cloud-index images derived from MSG HRV images. The procedure is shown in Figure 11.5. The most recent cloud-index image П0 at time t0 is compared with the preceding cloud – index image n_1 at time t_1 = t0—Dt, where Dt represents the time step between two consecutive images (Dt = 15 min for MSG images). Deriving cloud movement by comparing cloud structures in images П0 and n_1 is performed by assuming (1) constant pixel intensities for cloud structures in both images and (2) smooth wind fields, which usually exist at cloud heights. These assumptions allow for detecting cloud motion by matching the same cloud pattern in consecutive images (Figure 11.6).
Rectangular areas (target areas in Figure 11.6) in image n_1 around the origin of each motion vector (vector grid points) are compared to equally sized areas within their neighborhood (search area) to detect the advection of cloud patterns between these images (Figure 11.6).
The detection of cloud patterns from image n_1 in the subsequent image n0 is performed by minimizing the mean square pixel differences for these target areas, defined as
image n_i, image n, forecast image n+i
timet-15 min timet time t + 15 Min
FIGURE 11.5 Procedure for cloud-index image forecasts consisting of (1) detection of motion for existing cloud structures to evaluate the most recent cloud-index images; (2) application of the derived motion-vector field to the most recent cloud-index image to extrapolate the movement of cloud structures for the next hours; (3) smoothing procedure to reduce inaccuracies in the irra- diance forecasts.
FIGURE 11.6 Scheme to detect cloud motion and vector grid, target area, and search area for calculating CMVs. For each grid point, the cloud pattern in the target area of cloud-index image n_1 around this point is searched for in the cloud-index image For all target areas within the search area, the MSE is determined successively (a-c). The target area identified by the minimal MSE then defines the direction and length of the motion vector (d).
where d is the shift vector of all pixels x in the respective area. For each part of the search area, the MSE is calculated; the target area with minimal error is selected and defines the area’s motion vector. A more complex statistical method for the determination of CMV fields was also evaluated (Hammer et al., 1999). A Monte Carlo algorithm determines the probability of a transition between images through each possible motion-vector field, and it selects the most probable CMV for a cloud-motion forecast. The evaluation of this computationally demanding model showed no significant improvement regarding its applicability and resulting forecast accuracy.