Image Classification

Description of the Problem

Classifying large areas of lands can be problematic if for no other reason than their cost. While nothing can substitute for field observations, it can be very helpful to ascertain the general land cover classes in an area without, or prior to, entering the field. Land classification can be done within a GIS in two ways: supervised and unsupervised. Supervised classification, as presented here, allows a researcher to essentially train the GIS to look for patterns and from these patterns to classify its observations. The problem presented here is the need to classify part of the Black Water National Wildlife Refuge located in Cambridge, Maryland into six land cover classes (Forest, Cultivated Field, Barren Area, Developed/Impervious, Wetland, and Water) using raster data.

Strategies of Solving the Problem

The problem was approached by acquiring the relevant raster data from North Carolina State University.

Methods

After acquiring the data from North Carolina State University it was loaded into a GIS where the bands were then manipulated to depict an infrared image. From this image, training sample polygons were created. In order to increase accuracy of classification small polygons were used. This was in order to assure as homogenous a training sample as possible thereby increasing the discreetness of samples. After two training sample polygons were created and classified for each class an Interactive Supervised Classification was run. After this initial image classification was run it was determined that the classification could be improved by adding more sample training polygons into the classes that had issues (primarily Developed/Impervious and Water). After adding this second group of sample training polygons the Interactive Supervised Classification was run again showing improved results.

Map 1: Initial Training Sample Polygons
Map 1: Initial Training Sample Polygons
Map 2: Initial Image Classification
Map 2: Initial Image Classification
Map 3: Second Training Sample Polygons
Map 3: Second Training Sample Polygons
Map 4: Second Image Classification
Map 4: Second Image Classification

Discussion of Methods

The primary issue that was encountered was the initial training sample polygon sizes. Once it was realized that smaller, homogenous polygons would be favorable those were used and the procedure repeated.

Data Evaluation Procedure

Output consistency was checked by comparing it to the RGB aerial photography. The initial output left quite a bit to be desired due to the misclassification of several areas. Most notably, as shown on the maps, is the misclassification of some water areas as wetlands. Additionally there appeared to be complications with correctly classifying the areas of and around developed/improved land cover. In an attempt to correct this issue more training sample polygons were added to these classes in the hopes to differentiate them further from other classes. After creating a second output, as can be seen on the map, there was increased accuracy of classification most notably with water areas. However, there still appeared to be some issues in classification especially evident, still, around the developed/improved land cover even though these results did appear to improve from one output to the other.

Reflection and Ideas of Other Applications

The ability to classify land cover remotely, cheaply, and rapidly opens this application up to a variety of disciplines. Archaeology has been able to use this application to determine where sites might be visible. Supervised, or unsupervised as well for that matter, should not be the end of research. A GIS can only classify land cover from the training sample polygons that a user has input which has its own issues not least of which is user preference or perceived superiority of a sample area. In order to adequately classify land cover at least some field work must be done to observe actual land cover.

Advertisements