Automated Image Segmentation

How to turn a USGS map or satellite photos into vector data? A couple of links, and some questions.

Back when Adam Golding and I prototyped Probability Mapper we already had an algorithm that could give the probability based on distance, terrain, vegetation, and other factors. But we were working with raster images, so we only had distance. If you had vector layers for terrain & vegetation, you'd be set.

~Originally 9 June 2008

 

Example from Feature Analyst's Landcover Tutorial

But suppose you don't. How do you take a raster image (above left) and turn it into a set of vector polygons (above right)? It's too tedious to do by hand. We thought about making a classifier, but that sounded like a side project that would expand without end. A recent email exchange brought it to mind again, and I did some looking. It seems there are good solutions out there, if you can pay.

Some software

Definiens eCognition

On April 4, I attended a demo by ESRI at GMU. At one stage, they had to count the number of tailings cones (dirt piles) that had been excavated from a runway, to see if the volume of dirt matched the cut-and-fill operation, or if there was extra dirt suggesting underground facilities. Rather than manually count the thousands of piles, they used Definiens eCognition software to circle several and say "this is what I want". That trained a classifier, which went and found all the rest of them, and returned a vector layer for the piles.

Feature Analyst

Feature Analyst seems to be quite capable, using context and some pretty nifty machine learning. The examples and customers specifically mention Landcover classification. (See the example image above.)

I think the software is by Visual Learning Systems, though it is also listed under Leica Geosystems / Erdas. The software is available for Leica's ERDAS Imagine, and ESRI's ArcView and ArcGIS.

Berkeley ImgSeg

Programmers looking for a smaller(?) package may be interested in the targeted ad suggested by gMail during my previous email exchange: Berkeley ImgSeg. It appears to do the same thing: you train a classifier by pointing to "yes" and "no" areas, it learns, and then it classifies your image for you. The appear to have a Python API, which gets points in my book.

Bayesian Network Classifiers

Ed Wright, my GIS colleague at IET has developed probabilistic feature classifiers by embedding our company's Bayes Net engine into a major image processing package. So far as I know, the GIS company hasn't been convinced there is a market for this sort of thing, and has not released it. (Also, for all we know, Feature Analyst might use similar technology. The nice results it seems to be getting are a hallmark of Ed's multi-layer, context-based approach.)

Questions

  • Is anyone using this for Search & Rescue? Care to comment, esp. on "real-time" use in a mission?
  • Is adequate vector data available?
  • Is there a good open-source equivalent?

UPDATE: Answers as they come in

Bob Koester

Bob reminds me that the NLCD classifications are available directly, so for broad-strokes classification of 30m squares, we don't have to extract features: Charles: In regards to classify land. What I used in the book was NLCD which is available for free. It was taken from Landsat data, so in some cases it may be a bit dated. But for wilderness searches it is probabily ok. If I recall the resolution is a 30 meter square at best. Other coarser resolution are available. But I know I could look at the data easily using tools on the web. So I think that would be the quick and dirty.

http://edc.usgs.gov/products/landcover/nlcd.html

Plenty of info on the web.

It is one of the models I mention in the book, that should be used for predicting POA.

Jim Donovan

Jim writes: The Feature Analyst is one of the best out there for the ESRI platform. The company (VLS) got its start through a NASA SBIR. Naive Bayesian classifiers have been used lots in the remote sensing realm.

 

Author: ctwardy

Charles Twardy started the SARBayes project at Monash University in 2000. Work at Monash included SORAL, the Australian Lost Person Behavior Study, AGM-SAR, and Probability Mapper. At George Mason University, he added the MapScore project and related work. More generally, he works on evidence and inference with a special interest in causal models, Bayesian networks, and Bayesian search theory, especially the analysis and prediction of lost person behavior. From 2011-2015, Charles led the DAGGRE & SciCast combinatorial prediction market projects at George Mason University, and has recently joined NTVI Federal as a data scientist supporting the Defense Suicide Prevention Office. Charles received a Dual Ph.D. in History & Philosophy of Science and Cognitive Science from Indiana University, followed by a postdoc in machine learning at Monash.

Leave a Reply

Your email address will not be published. Required fields are marked *