xdownscale 1.0.2

A PyTorch-based tool to downscale spatiotemporal data

xdownscale

xdownscale is a Python package for super-resolution downscaling of gridded datasets using deep learning. It supports a wide range of applications, including satellite observations, reanalysis data, and climate model outputs. Built with PyTorch and xarray, it enables efficient mapping from coarse-to-fine-resolution grids in just a few lines of code.

---

Installation

To install from PyPI, we recommend using a conda environment

conda create -n xdownscale python=3.10
conda activate xdownscale
conda install -c conda-forge pytorch cudatoolkit=11.8 cudnn
pip install xdownscale

To install from source:

git clone https://github.com/manmeet3591/xdownscale.git
cd xdownscale
pip install .

Or install from a zipped archive:

unzip xdownscale_package.zip
cd xdownscale
pip install .

---

Usage

import xarray as xr
import numpy as np
from xdownscale import Downscaler

# Create dummy coarse-resolution input and fine-resolution target
x = np.random.rand(128, 128).astype(np.float32)
y = (x + np.random.normal(0, 0.01, size=x.shape)).astype(np.float32)

input_da = xr.DataArray(x, dims=["lat", "lon"])
target_da = xr.DataArray(y, dims=["lat", "long"])

# Initialize the downscaler
ds = Downscaler(input_da, target_da, model_name="fsrcnn")

# Predict high-resolution output
result = ds.predict(input_da)
result.plot()

Available models:
srcnn, fsrcnn, lapsr, carnm, falsra, falsrb, srresnet, carn, oisrrk2, mdsr, san, rcan, unet, dlgsanet, dpmn, safmn, dpt, distgssr, swin

---

Description

xdownscale performs patch-wise training using PyTorch’s DataLoader and returns predictions as xarray.DataArray objects. It is designed to work with any gridded dataset and provides a flexible interface for model selection, training, and inference.

---

Sample Data

Sample input and target data are provided in the data/ directory for testing and demonstrations.

---

Development

To extend or customize the package:

• Modify model architectures in xdownscale/model.py
• Add training logic in xdownscale/core.py
• Customize patch extraction and utilities in xdownscale/utils.py

---

License

This project is licensed under the MIT License.