### Geog 104 - Image Classification Exercise

Objective: Consider the notion of image classification... grouping similar pixels together and differentiating them from dissimilar ones based on the digital numbers that record spectral response.

Data: For this exercise we'll again use R0012846.JPG, taken from Chuck's ill-fated hexcopter on 13 March 2013. This image is 3,648 by 2,736 pixels and records red, green and blue intensities for each pixel. The camera's sensor recorded light in three spectral bands, nominally red, green and blue.

Software: Any image viewing program that will give you coordinates of pixels and show you a pixel's color in RGB space should do for this. I'd suggest GIMP , but Photoshop (Photo\$hop?), and even MS-Paint will do. [What happened to MacPaint ? Maybe Paintbrush will do? -Ed.]

One can think of a remotely sensed three-band image as a file of numbers representing the red, green, and blue 'coordinates' of each pixel in a three-dimensional "color" or Data Number (DN) space. Quite often, each pixel's response in each band is recorded as an 8-bit binary integer in the range [0..255].

Then, one might suspect that pixels that show the same stuff on the ground would be close together in the DN (color) space. I.e., things that look deep green should have color numbers that are similar to each other; things that look light gray should have numbers that are similar to each other and different from the green things. Similarly, brown pixels should cluster together in the DN space, away from the both green and the gray pixel clusters. In this exercies we'll try to see whether this idea works.

Task : Select, and name three kinds of landcover (classes of pixels) in the image, such as healthy grass, so-so grass, bare(ish) soil, concrete, palm fronds, ferns, the tops of the utility covers, people, shadows, you name it. For each class, pick 10 training pixels that exemplify the class and note down on the tables below the R, G, and B component DNs. Summarize the DNs for each class by computing the minimum, maximum and mean of the values in your training sample. And finally compare them by plotting the range and mean of each class in each band and commenting on the degree of overlap between your samples.

Vocabulary: Supervised Classification, Unsupervised classification, DN, training set, signature.

Landcover class:_______________________________
Sample R DN G DN B DN
1
2
3
4
5
6
7
8
9
10
Class min.
Class max.
Class mean

Landcover class:_______________________________
Sample R DN G DN B DN
1
2
3
4
5
6
7
8
9
10
Class min.
Class max.
Class mean

Landcover class:_______________________________
Sample R DN G DN B DN
1
2
3
4
5
6
7
8
9
10
Class min.
Class max.
Class mean