45 Computer Vision News Computer Vision News Climate Downscaling computationally expensive, and that’s where deep learning methods from computer vision can help. The models applied to superresolve images can also increase the resolution of climate model data!” Previous attempts at downscaling existed within the climate community but were limited in scope. Introducing deep learning architectures from computer vision significantly advances the downscaling task and allows for integrating physical constraints. “Sometimes we predict rain, for example, but then the neural networks violate the simple physical laws, like conservation of water mass,” Paula explains. “We developed a model to ensure that they’re physically correct, so we have climate predictions that make sense. Sometimes the models even predict negative masses, and we make sure that doesn’t happen.” Paula’s novel approach introduces a new layer at the end of a neural network to act as a constraint, slightly rescaling the output to make it physically accurate. The method can be applied to standard architectures, such as CNNs, GANs, transformer models, and normalizing flows, which are great at super-resolving data. “The first thing people try is to add another loss term to the loss function,” she reveals. “That’s very common, but it worked terribly. We did a lot of tuning and training, but it didn’t work out. That brought us to the new idea of including this final constraint layer, which worked very well. After some struggle with the so-called soft constraint, where you have a regularization term in your loss function, we then developed this constraint layer.”
RkJQdWJsaXNoZXIy NTc3NzU=