Figure 3 | Scientific Reports

Figure 3

From: Predicting clustered weather patterns: A test case for applications of convolutional neural networks to spatio-temporal climate data

Figure 3

Schematic of the up-sampling and down-sampling steps. Each daily full Z500 pattern, which is on a 66 × 97 latitude-longitude grid, is converted to a contour plot represented by a RGB image of size 342 × 243 pixels with 3 channels representing red, green, and blue. This up-sampled image is then down-sampled to an image of size 28 × 28 × 3 using bi-cubic interpolation and further standardized by subtracting the mean and dividing by the standard deviation of the pixel intensities. These images are the inputs to CNN for training or testing. The down-sampling step is used to remove redundant features at small scales from each sample. Trying to learn such small features, which are mostly random, can result in overfitting of the network. Note that rather than converting the data matrix into a RGB image, CNN could be applied directly to the data matrix, which we have found to yield the same accuracies. See Data and Methods for further discussions.

Back to article page