What is Deep Learning? A subset of Machine Learning - a Deep Learning Algorithm improves its computing prowess automatically as it interacts with data i.e. it doesn't require explicit programming to do so.
Recollect how, in my previous post, Forest-based Machine Learning Algorithms were deployed to Classify Radar Imagery to identify Deforested Surfaces, Classify Optical Imagery to identifying Crop Types, and Classify the variables influencing Voter Turnout during Elections by feeding 'supervised observations' for the algorithm to 'train / learn' from, based on which it predicted the desired output with considerably high accuracy.
Section Hyperlinks below:
Deep Learning (DL) Algorithms use a method similar to how our brain functions - Neural Networks - to arrive at an accurate answer / decision when faced with multiple options / choices. The simplest example which I can think of where Deep Learning Algorithms are practically used are 'Search Results' - the likes of Amazon and Netflix use such algorithms to recommend content you are likely to appreciate based on your historical searches - which act as 'Supervised Observations' for the algorithm to train upon. Think how often the next song / video in your playlist recommended by Spotify / YouTube is so close to what you'd love to listen / view at that very moment! - this demonstrates the utility of such algorithms
The video explainer below effectively highlights how 'Artificial Neural Networks' - the fundamental operating mechanism of Deep Learning Algorithms - work:
Video 1: Artificial Neural Networks Explained. Source: Esri's Spatial Data Science MOOC
You may choose to read this article which outlines 20+ Applications of Deep Learning Algorithms.
In Geospatial workflows, Deep Learning Algorithms are particularly used in Imagery Analytics for -
a) Object Detection (identifying imprints),
b) Instance Segmentation (identifying boundary of the object detected),
c) Image Classification (using predefined rules to identify whether an object is X or Y or Z), and
d) Pixel Classification (a form of semantic segmentation - identifying whether a pixel is from a desert, an ocean, a forested area, and so on) purposes.
In a subsequent post, I have used a Deep Learning Model to classify Power Lines. Check it out.
The more quality supervised observations the algorithm is trained upon, besides other factors, the more accurate its output tends to become. Advanced Mapping Platforms now integrate ready-to-use Deep Learning Models to aid multiple workflows - a couple of which I'll highlight in this post.
Workflow 1: Detecting and Demarcating Building Footprint
One of the two geospatial workflows involving the practical usage of Deep Learning Algorithms covered in this post is to extract Building Footprints from Optical Satellite Imagery.
A Building Footprint, as the name suggests, is an 'imprint' of the buildings within a predefined geographic extent. By accurately detecting and extracting the footprint, one can demarcate the building outline as well as estimate certain dimensional characteristics about the object - such as the Area occupied by the Building. A 'Building Footprint' geospatial layer - comprising of multiple building imprints within a geographic extent - acts as an important dataset in several workflows - be it Urban Planning and Property Insurance or Public Transportation and Physical Security.
In my previous post / video, I had demonstrated a workflow involving the estimation of Rooftop Solar Power Generation potential in a locality which involved the usage of a Building Footprint geospatial data layer. Using Solar Siting tools over this layer allowed me to estimate how much sunlight each roof will receive annually and, as a result, estimate its annual Power generating potential if rooftop solar panels were to be installed.
Digitizing Building imprints manually is a tedious process and susceptible to errors. Thanks to advances in Technology - high resolution Satellite Imagery, increase in Computing speeds and the use of Object Detection DL Algorithms, performing this workflow is now highly automated, much quicker and more accurate than traditional methods deployed not so long ago.
The ready-to-use Deep Learning model used in this demonstration was developed by Esri - the world's leading GIS platform developer - and it was trained on a large quantity of Imagery datasets with 30-60 cm spatial resolution over several representative regions within the USA. While the Algorithm works best for the detection of buildings with similar exterior characteristics i.e. Buildings within USA as it was trained on them, however, it also fares reasonably well on other developed countries too - which tend to have similar Building exteriors as found in USA.
The image below is hyperlinked to an engaging 'Storymap' where you'll be able to witness the Deep Learning Algorithm output for this workflow.
Figure 2: Image from a Storymap which contains samples of Building Footprint extracted by Esri's ready-to-use Deep Learning Model over Sweden
For my demonstration, I have decided to test Esri's Deep Learning model on a 30 cm spatial resolution Optical Imagery from 2009 over a cross-section near the Barajas Airport in Madrid, Spain which I accessed from European Space Imaging.
Upon running the Model with conservative parameters, the output generated is depicted below-
(The sliders below are best viewed on a PC)
Output 1: Building Footprint Extraction using Esri's ready-to-use Deep Learning Model
Output 2: Building Footprint Extraction using Esri's ready-to-use Deep Learning Model
The Algorithm used its trained knowledge about Building Types in the USA and proceeded to detect and demarcate similar looking Objects from this Imagery over Madrid (masked in green).
It would be fair to assume that the extracted Footprint would have been more precise had I tested different, more relevant parameters. Also, I did not have access to the 'Regularize Footprint' geoprocessing tool which would have allowed me to extract more accurate Building outlines.
In another post, I've been able to do so - and it also contains a detailed video about the workflow within. The second workflow covered below also contains a video demonstration.
That being said, from the Model's perspective, as its developers feed in more, diverse supervised Building outline observations and also provide feedback / finetune the existing output generated (in terms of false positives and false negatives), it will evolve and become better at what it does.
Workflow 2: Detecting and Demarcating Swimming Pools
The second workflow covered in this post involves the detection and demarcation of Swimming Pools in the city of Redlands in California, USA. This information will be useful for Tax Assessors as the presence of Swimming Pools within a property enhances its value and resultingly, increases the applicable Property Tax. You can imagine the utility of an automated detection and demarcation workflow for this use case - the assessors are currently only able to obtain this information / update their records by conducting infrequent manual surveys.
Methodology-wise, the workflow is very similar to that of detecting Building Footprint - we feed in supervised observations to train the DL Model, input the Processing Parameters, allow the Model to ingest and validate it, and finally deploy it to extract Swimming Pools from the specified geographic extent. A video demonstration of this workflow can be seen below-
The video is best viewed in YouTube's HD setting either on a Desktop or using Landscape mode on Mobile Phones. You can reduce / increase the play speed as per your viewing preference.
Video: Demonstration for detecting Swimming Pools in the city of Redlands in California, USA using Esri's Deep Learning Model
Isn't this fascinating?
The potential of Deep Learning Algorithms are limitless and I wouldn't blame you were you to even find it scary - even at this preliminary stage, the algorithms are able to defeat the best humans in business - watch this brilliant documentary - AlphaGo - which involves the use of Google's powerful DeepMind AI algorithm in a Human vs Computer contest for a popular strategy game called Go.
ABOUT US
Intelloc Mapping Services | Mapmyops.com is based in Kolkata, India and engages in providing Mapping solutions that can be integrated with Operations Planning, Design and Audit workflows. These include but are not limited to - Drone Services, Subsurface Mapping Services, Location Analytics & App Development, Supply Chain Services, Remote Sensing Services and Wastewater Treatment. The services can be rendered pan-India, some even globally, and will aid an organization to meet its stated objectives especially pertaining to Operational Excellence, Cost Reduction, Sustainability and Growth.
Broadly, our area of expertise can be split into two categories - Geographic Mapping and Operations Mapping. The Infographic below highlights our capabilities.
Our 'Mapping for Operations'-themed workflow demonstrations can be accessed from the firm's Website / YouTube Channel and an overview can be obtained from this flyer. Happy to address queries and respond to documented requirements. Custom Demonstration, Training & Trials are facilitated only on a paid-basis. Looking forward to being of service.
Regards,
Much Thanks to Esri & European Space Imaging for the training material