geodl: An R package for geospatial deep learning semantic segmentation using torch and terra.

Area of Science:

• Geospatial data science
• Earth observation
• Geophysical data analysis

Background:

• Deep learning (DL) methods, particularly Convolutional Neural Networks (CNNs), excel at analyzing complex spatial data.
• Existing R packages facilitate geospatial data processing but lack integrated DL capabilities.
• Implementing DL for geospatial tasks often requires complex Python/PyTorch environments.

Purpose of the Study:

• Introduce the geodl R package for seamless geospatial deep learning in R.
• Provide a simplified workflow for pixel-level classification of Earth science data.
• Enable R users to leverage advanced DL techniques without external dependencies.

Main Methods:

• Leverages the torch package for DL implementation within R, integrating C++.
• Utilizes the terra package for efficient processing of large-scale, multi-band raster data.
• Incorporates the luz package for training loop management and custom DL model components.

Main Results:

• geodl facilitates the creation of raster masks and data subsets for DL model training.
• The package supports UNet-based models with customizable modules and loss functions.
• Includes standard remote sensing assessment metrics and enables predictions on extensive spatial datasets.

Conclusions:

• geodl significantly lowers the barrier to entry for applying deep learning to geospatial data within the R environment.
• The package streamlines the entire workflow from data preprocessing to model assessment.
• Future development aims to expand functionality and support for diverse geospatial DL applications.