|
Explore Teaching Examples | Provide Feedback

Image Classification

Based on materials by Dr. Paul Cote, Graduate School of Design, Harvard University

Starting Point pages compiled by Dr. Brian C. Welch, St. Olaf College

Author Profile

This material is replicated on a number of sites as part of the SERC Pedagogic Service Project

Summary

The purpose of this assignment is to get some hands-on experience with the fundamentals of image classification. Multi-Band images of the Earth's surface are becoming a very important source of information about land cover and land use. Because satellites beam back information every day, this imagery can be a terrific source of very current information or historic (perhaps back to 1972). Of course, subtracting the historic from the current can provide an estimate of change in the landscape, provided one can be sure that the classification of the images yields consistent results.

Used this activity? Share your experiences and modifications

Learning Goals

Students who complete this exercise will learn the following:

  • what are "multi-spectral" or "multi-band" images?
  • how landscapes change over time - at least over the time-span of satellite imagery, particularly due to human influences
  • how to train an image analysis program to identify types of land surfaces based on the spectral (color) content of the recorded image
  • how to move raster (image) files between ArcView and MultiSpec
  • to use MultiSpec image analysis program

Context for Use

This exercise can be used as a lab or take-home exercise if the students have access to the Internet. It can be used to supplement discussions about satellite imagery, human impact on the environment, GIS, geomorphology, etc. Ideas for adapting the exercise to other topics are provided in the Teaching Notes.

The full exercise will take at least 3 hours, but could be shortened or lengthened to fit nearly any time-frame

The exercise requires the use of MultiSpec, a free image analysis program from Purdue University. MultiSpec runs on Windows and Macintosh computers. An Internet connection is needed to download the data sets unless they are provided to the students by the instructor. The use of ESRI's ArcView by the students is expected, but could be bypassed in ways described in the Teaching Notes below. Basic ArcView skills such as georeferencing a raster file and knowledge of basic Raster Calculator functions in the Spatial Analyst extension are assumed.

Teaching Materials

This exercise focuses on landscape changes on Cape Cod, Massachusetts. The exercise could be modified to work with any pair of geolocated multi-spectral images that show change over time of a specific feature.

  • The original exercise was by Dr. Cote at the Harvard Graduate School of Design. The contents of the original exercise may change over time as it is modified and improved.
  • A hand-out (Microsoft Word 1MB Mar28 05) of the exercise as an MS Word document. This file can be edited to apply to the specific classroom environment where it will be used. This file may not match exactly the contents of the online version at Harvard.

Teaching Notes and Tips

Image classification - background

Image classification is a means to convert spectral raster data into a finite set of classifications that represent the surface types seen in the imagery. These may be used to identify vegetation types, anthropogenic structures, mineral resources, or transient changes in any of these properties. Additionally, the classified raster image can be converted to vector features (e.g. polygons) in order to compare with other data sets or to calculate spatial attributes (e.g. area, perimeter).

Image classification is conducted in three modes: supervised, unsupervised, and hybrid. These are described briefly in Ch. 8 of the textbook. In general, a supervised classification requires the manual identification of known surface features within the imagery and then using a statistical package to determine the spectral signature of the identified feature. The "spectral fingerprints" of the identified features are then used to classify the rest of the image. An unsupervised classification scheme uses spatial statistics (e.g. the ISODATA algorithm) to classify the image into a predetermined number of categories (classes). These classes are statistically significant within the imagery, but may not represent actual surface features of interest. Hybrid classification uses both techniques to make the process more efficient and accurate.

MultiSpec

MultiSpec is a freeware program from Purdue University that is a very effective image-analysis package. It runs on Windows and Macintosh computers. Image analysis is very memory-intensive, so the program will run faster and more effectively on machines with more RAM available.

Students will need some time to become familiar with the program. Creating the training polygons is a little tricky and can be frustrating since erroneous polygons cannot be fixed, but must be completely removed and redrawn. Students will also find the apparent arbitrary and non-repeatable nature of the training polygons frustrating (hence the development of automated classification schemes), but this is part of the exercise. Image classification and analysis is only as good as the method used to describe the spectral nature of each type of land surface (e.g. trees vs. water vs. buildings, etc.). In the end the students will find that errors in the polygons usually result in only small differences in the results - as long as they are reasonably careful.

MultiSpec can be downloaded individually (see Resources below), or with the data sets for this exercise as part of the Zip file.

Data sources for the exercise

The MultiSpec program, images and other data files are available in a 293 MB zipped (compressed) file. This is a very large file and may take a considerable time to download, so it should be done by the instructor before class. The file can then be copied to the classroom computers, a local server, or a CD to give the students access.

Unzip the file using standard decompression software (e.g. WinZip) into a new directory to keep the files organized. MultiSpec is contained as an executable within the Zip file. Once it is decompressed, the "multispec" directory can be moved to a different location so that it is available for use with other image files.

  • Download files from SERC (this site)
  • Download files from Harvard (original site)(http://www.gsd.harvard.edu/geo/manual/image_class/gsd/ic_lab.zip). Note: The filenames may change if the files are updated by Dr. Cote. Go to the original exercise site at Harvard if the data link is broken to find the new data file link.

Bypass the use of ArcView by the students

The image files in the data set are geolocated, but each band is in a separate file. MultiSpec requires a single multi-band file, so ArcView or some other image-processing tool (e.g. ENVI) is needed to create files for MultiSpec. One way to shorten the exercise or to focus on the image analysis rather than file manipulation is to create MultiSpec-ready multi-band files for the students .

Calculating the area of each type of land surface in the classified image requires loading the classified image file into ArcView, converting the image to polygons, and then creating a new database field and script to calculate the area for each landscape type. While these are all useful GIS skills, the time and complexity involved may detract from the geoscience goals of the exercise (looking at changes in the Earth's surface features). Therefore, once the students create their own classifications and turn them in, the instructor can calculate area for a single representative sample file for each image date. The calculation results are then sent to the students to complete their write-up and discuss the quantified land surface changes that they mapped in the imagery.

The instructions for both of these options are spelled out in the exercise hand-out and require the use of MultiSpec and ArcView 8.x or ArcView 9.x and the Spatial Analyst extension.

Other geoscience applications of the exercise

The methods described in this exercise can be used for nearly any pair of geolocated images that show change over time. This means that the exercise could cover a wide variety of geoscience topics and regions that can be viewed through remote sensing data.

  • Use Measuring Distance and Area in Satellite Images (College Level) exercise ideas, but substitute MultiSpec (instead of ImageJ) as a more robust GIS tool to classify the entire image and then calculate polygon areas in ArcView.
  • Climate - Sea ice, icebergs (e.g. C-16 from Ross Sea), glacier retreat, desertification, lake evaporation, etc.
  • Geomorphology/Volcanology - Mount St. Helens pre-/post-1980 or pre-/post-2004-05 eruption.
  • Geophysics - changes in magnetic declination or field strentgth
  • Hydrology - pre-/post-flood imagery (e.g. 1993 Mississippi flood).
  • Solar System - changes seen in images from Mars or other objects in the solar system. It may be difficult to obtain spatially referenced images that can be recognized by ArcView and MultiSpec.

Assessment

This exercise has the students process and analyze satellite images using a simple piece of software. The exercise hand-out specifies a number of map and image products that the students should produce. The results of these analyses can become part of a write-up describing their experiences, observations, and conclusions regarding the changes seen in the pair of images. The quality of their results (e.g. accuracy of the training polygons and resulting classification) is a measure of how well the students comprehend the methods and concepts.

References and Resources