Code-Free Deep Learning for Geospatial Applications

Abstract

In the geosciences, deep learning (DL) has been used to develop state-of-the-art methods for weather and climate prediction, seismic signal processing, and remote sensing analysis, among other tasks. However, developing a DL model demands an understanding of advanced programming and statistical techniques as well as domain knowledge of the desired task. Additionally, popular DL libraries do not currently have sufficient functionality for use in geoscience, which often works with large volumes of high-dimensional data in nonstandard file types, such as vector files or hyperspectral images. We have developed a novel, no-code DL workflow for geoscience applications that guides users in training a custom model and using it to produce classified geospatial data ready for input into any GIS software. It is completely open-source and distributed freely via Google Colaboratory. Our model is based on the UNet architecture, a deep convolutional neural network proficient at geospatial pixel-wise classification tasks due to the ability to maintain relative spatial information about the input data. Customization of the provided model is handled by interactive widgets. Geospatial inputs are stacked as image channels, allowing a model to learn from a variable number of inputs instead of only RGB color bands. For example, a trained model could analyze hyperspectral imagery alongside LiDAR to learn identifying features from both inputs. The workflow addresses the class imbalance problem prevalent in geospatial datasets through optional oversampling during data preprocessing and the use of focal loss during training. We demonstrate the utility of the workflow by implementing it on three different satellite and airborne LiDAR datasets. Due to the workflow's modularity, custom geospatial functionality, and open-source format, it has the potential to assist with a broad range of geospatial classification tasks.