More Classification Work

The morning today (Saturday the 22nd) was spent continuing work in eCognition working towards improving my classification of kuku’i for the final project. I kept rebuilding feature extraction parameters working to identify the difficult to find kuku’i in the landscape. This is frustrating because it is clear to anyone looking at the screen as to which is kuku’i and which is other vegetation. However, when utilizing the parameters in eCog, the pixel values are averaged across the polygon of interest, which makes things like certain sections of pasture or high level vegetation blend in with kuku’i. This certainty decreases the overall accuracy of this object based method. There are sufficient variables built into the workflow of the software that I am certain that it is possible to isolate the kuku’i near perfectly, however when you must decide your value for scale parameter, shape, compactness, layer weights, and feature extraction methods, there are an innumerable quantity of combinations which all result in very different products. After working with this software for the past few days, I believe I am getting better at deciding these parameters but feel like it would likely take multiple years to gain a true mastery of this program. This afternoon around 4 we all went around and talked about our project progress and plans. I really enjoy what I am working with and studying however I hope to produce a useful product and must get past all the different small associated problems in order to build a correct classification. I am worried that when I quantify my final accuracy of the WV2 classification that it will not be what I wanted, and am continuing to work towards improving this classification day by day.