Nrdly
Get Nrdly Free Trial Built with Nrdly

Course 3. Deep Learning for Mapping

Introduction

The rapid advancement in hardware (such as graphics processing unit and tensor processing unit) coupled with developments in convolutional neural networks and the availability of very high-resolution (VHR) images has led to an interest in applying deep learning techniques in the geospatial field. Deep learning is an ensemble of techniques and algorithms that exploit neural networks to solve computer vision or natural language problems. Deep learning is a subset of machine learning methods, which comprises object detection, semantic segmentation, and instance segmentation techniques and algorithms.

Deep learning techniques and algorithms for image interpretation can be grouped into fully convolutional and region-based approaches. Fully convolutional methods use an encoder-decoder architecture such as semantic pixel-wise segmentation (SegNet), U-Net, and SharpMask. In contrast, region-based processes use feature extraction based on a stack of convolutional neural networks (CNNs). These techniques include Mask Region-based Convolutional Neural Networks (Mask RCNN), Pyramid Scene Parsing Network (PSP Net), and Deep Labeling for Semantic Image Segmentation (DeepLab). While deep learning algorithms have successfully solved general computer vision problems, their implementation for practical geospatial applications remains challenging. This is due to deep learning application specifications, which require the appropriate hardware (graphics processing unit), different and complex software libraries, and adequately labeled training datasets.

This course provides guidelines on implementing deep learning-based semantic segmentation to detect or map urban features such as building footprints and roads. We are going to use VHR imagery for mapping the urban elements. This course consists of two labs. Lab 1 will focus on instance segmentation using Mask-RCNN on a local machine (CPU or GPU), while lab two will focus on instance segmentation using Mask-RCNN on a Google Colab.