A Comparison Of Pixel Based And Object Based Image Analysis With Machine Learning Algorithms For Land Use/ Land Cover Classification

ABSTRACT

High resolution imageries can help understanding dynamic and heterogeneous landscapes much better. Many earth observatory satellites have been launched since advent of this technology. The importance of classification as an application of Remote Sensing to understand the pattern and characteristics of the landscape is much acknowledged. The overall classification accuracies for both the approaches was more or less same but the pixel-based approach took lesser time as the number of variables used were comparatively lesser. There has been no coherent work describing the difference between pixel based and object based classification stating the influence of sampling strategy and contextual information/segment size over the overall and within class accuracy. Moreover, based on the overall accuracies and per class accuracy the object based classification was agreed to have better results. The imbalanced training data resulted in reduction of accuracy for all the three classifiers. Mixed training data for object based classification also yielded higher accuracies as compared to pure training data. The main emphasis in this study is given to study the differences between various machine learning algorithms with the pixel based and object based classification techniques. Furthermore, the objectives can be summarized as to differentiate between the different classifiers based on sampling strategy (size and purity of the sample) and to assess the influence of the attributes of the training data on different class accuracy and overall accuracy.

FULL TEXT