Image Classification

11/13/14

Problem:

Managers of Black Water National Wildlife Refuge in Maryland, USA, have requested classification of the land cover for the waterfowl sanctuary in order to develop an appropriate land management plan. They have provided an aerial photograph of the refuge captured in August, 2010. The image contains four bands, true color and color infrared at 1-foot resolution. The images were delivered in meters, but due to the prevalence of acreage as the preferred unit to express land cover areas in the USA, the results will be presented in acres. The data used for these analyses were acquired from Dr. Perver Baran and Dr. Stacy Nelson at NCSU. The data included a single four band color image for the Black Water National Wildlife Refuge. The coordinate system used for these data was NAD_1983_StatePlane_ Maryland_FIPS_1900, meters. The image contains six main land cover classes including water, wetland, barren land, forest, cultivated fields, and impervious surfaces.

Analysis Procedure:

The general method used to perform these analyses involved first downloading the data and importing it onto ArcMap. The image required conversion from true-color to color-infrared within ArcMap before image analysis operations were performed. Image analysis should be performed in two stages, initial and modified, using pixel classification training sets generated manually. The initial stage broadly identifies trends in color classes, followed by modified stage to fine tune the pixel assignments. In the initial stage, three polygons within areas of each of the six land cover types should provide an adequate foundation upon which to run the Interactive Supervised Classification tool. In the modified stage, the number of training polygons will be increased, particularly in those areas that were not appropriately assigned in the initial stage, followed by another instance of the Interactive Supervised Classification tool. The resulting classification outputs from the tool will be exported to GRID layers and added to the map document, where the areas of each of the land cover classes can be calculated in acres.

Figure 1. Workflow diagram for Image Classification

Results:

Based on the initial training polygons (Figure 2), land classification compared to the original image (Figure 3) showed area classified as impervious surface in the water and wetland areas, and a substantial amount of crop classified as forest. Barren and shadows were also considered wetland. The forest was relatively well characterized. After refining the training polygons (Figure 4), the modified classification compared to the original image (Figure 5) improved. Most of the wetland areas became wetlands with some confusion with shadows and water, and to some degree barren areas. The crop land areas were largely identified properly, but there remained some confusion with forest, and visa versa.

Figure 2. Initial set of training polygons

Figure 3. Image Classification performed with an initial set of training polygons

Figure 4. Modified set of training sample polygons

Figure 5. Image classification performed with additional training polygons

Application and Reflection:

There is a place for using color in image analysis, especially as an initial stage to identifying the characteristics that are of interest, and it can be very useful for identifying objects that have distinctly different spectral signatures. But, it is imperfect. The color green (or, any color, for that matter) has thousands of shades. If the tolerances in the image processing tool can be set for each shade, it may be possible to pick colors that tease apart the objects of similar color, but the training set will need to include all these various shades in order to perform properly. Assuming one was able to generate such a training set for one image, when shifting from image to image, the qualities might vary (cloud cover, shadowing, water vapor differences, etc.), making the training set unable to conduct the same analyses on different images. Creating that training set could take an inhibitive amount of time in a process that manually draws training polygons. The image we worked with for this assignment added a fourth, color infrared band of information which helped substantially, yet there remained similarities in some of the features (crop vs. forest).

In agricultural applications of remote sensing, for example determining locations in a field under nutrient stress, the image processing activities would benefit from a technique that will reliably pick out spots that have a range of subtle color variation due to fertilizer availability. More spectral bands may be able to help even further, particularly if tunable to the wavebands most impacted by the treatment. For a field under nutrient stress, a multi-spectral image output with wavebands in the blue, green, red, and red-edge regions would be useful. For broad acre areas, these data might be obtained from satellite imagery or manned, aerial vehicles, and for information specific to a particular field, these data might be obtained by tractor-mounted imaging systems. With these information, tools that extract vegetation indices, for example the ArcMap NDVI tool, can be used to analyze the images.