A Python package to fusion LR and HR imagery 🚀

---

GitHub: https://github.com/IPL-UV/sathybrid 🌐

PyPI: https://pypi.org/project/sathybrid/ 🛠️

---

Overview 📊

Sathybrid is a Python package designed for fusing low-resolution (LR) and high-resolution (HR) satellite imagery. It allows users to combine spatial and spectral details from different sources, enhancing imagery through methods like Fourier transform fusion and interpolation techniques. The package supports multiple downsampling methods, and it also includes features like image denoising, histogram matching, and kernel-based blurring.

Key features ✨

• Image fusion: Combines LR and HR images using Fourier-based interpolation and various resampling techniques (e.g., Lanczos, cubic). 🌍
• Denoising with SwinIR: Utilizes the SwinIR model for efficient image denoising prior to fusion. 🧹
• Custom kernels: Apply customizable kernels such as Lanczos and cubic for image processing. 🎛️
• Fourier filtering: Supports multiple Fourier filters (ideal, Butterworth, Gaussian) for frequency-domain fusion. ⚙️
• Efficient resampling: Resizes images with support for different interpolation methods like Lanczos and cubic. 📐

Installation ⚙️

Install the latest version from PyPI:

pip install sathybrid

How to use 🛠️

Basic fusion of HR and LR imagery 🛰️

Load libraries

import sathybrid
import pathlib

Define the paths for high-resolution (HR) and low-resolution (LR) images

PATH = pathlib.Path("/home/cesar/demo/NA5120_E1186N0724/")
HRfile = PATH / "naip" / "m_3812243_nw_10_060_20220524.tif"
LRfile = PATH / "s2" / "s2_image.tif"
OUTfile = PATH / "fusion.tif"

Perform the image fusion

sathybrid.image_fusion(
    hr_file=HRfile,
    lr_file=LRfile,
    output_file=OUTfile,
    scale_factor=8,
    hr_bands=[1, 2, 3],
    hr_normalization=255,
    lr_bands=[3, 2, 1],
    lr_normalization=10_000,
    upsampling_method="lanczos3",
    fourier=True,
    fourier_params={"method": "butterworth", "order": 6, "sharpness": 3},
    denoise=True,
)

Finding the most similar LR image 🖥️

Load libraries

from sathybrid.utils import find_similar_lr

Identify the most similar LR image to the HR image based on FFT-L1 similarity

data_stats = find_similar_lr(
    hr_file=HRfile,
    lr_folder=PATH / "s2",
    hr_bands=[1, 2, 3],
    lr_bands=[3, 2, 1],
    downsampling_method="lanczos3",
    method="fft_l1"
)

Get the most similar LR image

best_lr_image = data_stats.iloc[0]["lr_img"]
print(f"Most similar LR image: {best_lr_image}")

Applying custom kernels🎛️

Load libraries

from sathybrid.blur import apply_kernel_to_image
import torch

Load a sample image (in PyTorch tensor format)

image = torch.randn(1, 3, 256, 256).cuda()

Apply a triangle kernel

smoothed_image = apply_kernel_to_image(image, kernel_size=5, method="triangle")

Advanced usage ⚙️

Image denoising with SwinIR 🧹

Load libraries

from sathybrid.main import setup_denoiser, image_denoise

Initialize the SwinIR denoiser model

denoiser = setup_denoiser()

Load the image to be denoised

image_tensor = torch.randn(1, 3, 256, 256)

Denoise the image

denoised_image = image_denoise(image_tensor, denoiser)

Image Fusion with fourier transform 🧮

Load libraries

from sathybrid.main import image_fusion

Define paths for HR, LR, and output

HRfile = "/path/to/hr_image.tif"
LRfile = "/path/to/lr_image.tif"
OUTfile = "/path/to/output/fusion.tif"

Perform fusion with fourier filtering

hybrid_image, error = image_fusion(
    hr_file=HRfile,
    lr_file=LRfile,
    output_file=OUTfile,
    scale_factor=4,
    fourier=True,
    fourier_params={"method": "gaussian", "sharpness": 2.5},
    denoise=False,
)

Supported features and filters ✨

• Resampling methods:

  • lanczos3, lanczos5, lanczos7, cubic, nearest.

• Fourier filters:

  • ideal, butterworth, gaussian, sigmoid