How to extract surface from air photo avoiding trees and buildings?

Question

I am working with Drone air images. These are geotiff images and clearly portray trees and buildings. I need to first classify the image according to land cover distribution, then need to remove the trees, grasses and buildings to get the original terrain model. I have another image of same area with elevation values and I need to reduce the elevation of objects over the surface.

I used ArcGIS 10.2 Spatial Analyst>Image Classification. I selected the Training Areas against my classification requirement and the generate signature file using automated toolset. I used this signature file for supervised classification and used Maximum likelihood for my classification.

The result of classification was awful. I found the big trees are classified as both grass land and trees as well as crops field also classified as big trees.

The image is RGB and it shows similar reflectance of big trees, crop fields and grass land. So, it is difficult for me to classify different classes separately to extract actual surface reducing tree height or crop height.

I am wondering if anybody has any suggestions regarding this issue.

Answer 1

I wanted to write a comment, but i don´t have points enough. I think that what you need is to do a classification by texture. Last week i was on a seminar where the aim was to classify images of high resolution with texture and variograms (geostatistics).

you can read this:

• http://www.sciencedirect.com/science/article/pii/S0098300499001181
• atkinson & lewis. Geostatistical classification for remote sensing: an introduction

i hope this helps you!

Answer 2

Do you have access to the point cloud from the imagery? In mosaicing the images from the drone, depending on software, you can export a 3D point cloud. You then can use LASTools to classify the ground points and then convert to a DEM.

Answer 3

Three band images are generally not sufficient for high quality land cover classifications. Usually at least near infrared band is required. When I was classifying one image that had four bands (r,g,b,nir) I also calculated NDVI and included it in classification. As you probably don't have nir band you could add more information for the classification using other indexes that require two of three bands that you have at your disposal. In this article there is a list of some that you might use.