US 12,394,071 B2
Temporal interpolation of precipitation
Takao Moriyama, Yokohama (JP); Michiaki Tatsubori, Oiso (JP); Tatsuya Ishikawa, Chuo-ku (JP); and Paolo Fraccaro, Warrington (GB)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Dec. 21, 2021, as Appl. No. 17/557,490.
Prior Publication US 2023/0196594 A1, Jun. 22, 2023
Int. Cl. G06T 7/269 (2017.01)
CPC G06T 7/269 (2017.01) [G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30192 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A computer-implemented method for training temporal precipitation interpolation models, comprising:
receiving, by one or more processors, an initial image, a first intermediate image, and a final image showing precipitation in a region, wherein the initial image is captured at an initial time, the first intermediate image is captured at a first target time after the initial time, and the final image is captured at a final time after the first target time;
computing a first preliminary forward optical flow vector field from the initial image to the first target time, and a first preliminary backward optical flow vector field from the final image to the first target time using a first neural network;
computing a first refined forward optical flow vector field and a first refined backward optical flow vector field for the first target time by inputting (i) the initial image, (ii) the final image, (iii) the first preliminary forward optical flow vector field, (iv) the first preliminary backward optical flow vector field, and (v) a terrain factor into a second neural network, wherein the terrain factor comprises a rain motion pressure feature calculated via an element-wise negative inner-product of the first preliminary forward optical flow vector field and a topographic gradient vector field of a topographical elevation map; and
computing backpropagation losses to train the first neural network and the second neural network by comparing the first intermediate image to an interpolated frame calculated using the first refined forward optical flow vector field and the first refined backward optical flow vector field.