carlos-gg / dl4ds

Deep Learning for empirical DownScaling. Python package with state-of-the-art and novel deep learning algorithms for empirical/statistical downscaling of gridded data

Deep Learning for empirical DownScaling

DL4DS (Deep Learning for empirical DownScaling) is a Python package that implements state-of-the-art and novel deep learning algorithms for empirical downscaling of gridded Earth science data.

The general architecture of DL4DS is shown on the image below. A low-resolution gridded dataset can be downscaled, with the help of (an arbitrary number of) auxiliary predictor and static variables, and a high-resolution reference dataset. The mapping between the low- and high-resolution data is learned with either a supervised or a conditional generative adversarial DL model.

The training can be done from explicit pairs of high- and low-resolution samples (MOS-style, e.g., high-res observations and low-res numerical weather prediction model output) or only with a HR dataset (PerfectProg-style, e.g., high-res observations or high-res model output).

A wide variety of network architectures have been implemented in DL4DS. The main modelling approaches can be combined into many different architectures:

Downscaling type	Training (loss type)	Sample type	Backbone section	Upsampling method
MOS (explicit pairs of HR and LR data)	Supervised (non-adversarial)	Spatial	Plain convolutional	Pre-upsampling via interpolation
PerfectProg (implicit pairs, only HR data)	Conditional Adversarial	Spatio-temporal	Residual	Post-upsampling via sub-pixel convolution
			Dense	Post-upsampling via resize convolution
			Unet (PIN, Spatial samples)	Post-upsampling via deconvolution
			Convnext (Spatial samples)

In DL4DS, we implement a channel attention mechanism to exploit inter-channel relationship of features by providing a weight for each channel in order to enhance those that contribute the most to the optimizaiton and learning process. Aditionally, a Localized Convolutional Block (LCB) is located in the output module of the networks in DL4DS. With the LCB we learn location-specific information via a locally connected layer with biases.

DL4DS is built on top of Tensorflow/Keras and supports distributed GPU training (data parallelism) thanks to Horovod.

API documentation

Check out the API documentation here.

Installation

pip install dl4ds

Example notebooks

A first Colab notebook can be found in the notebooks folder. Click the badge at the top to open the notebook on Google Colab.