SPATIAL OBJECT BASED REMOTELY SENSED IMAGE CLASSIFICATION TECHNIQUE FOR TEA PLANTATION MAPPING IN SRI LANKA

摘要

This study presents a spatial object based remotely sensed image classification technique (SORSICT) to identify the tea plantation land-use from satellite imagery. In Geographic Information Systems, supervised learning is used to classify remotely sensed imagery data to develop land-use maps. The traditional supervised classifiers used for this process tend to generate irrelevant classes that subjected to use the post-classification methods for correcting the errors of the land-use classes. This research found a solution for the problem of how do the current image classification algorithms be refined to generate accurate land use thematic maps for the tea plantation in Sri Lanka. The new approach is a refined supervised classification which uses the conventional minimum distance decision rule. It incorporates a spatial object based threshold scheme to limit the insertion of the irrelevant land-use types in a particular region. That spatial threshold scheme is a vector-polygon layer generated by a Geographic Information System using the prior knowledge about the land-use pattern of the particular region. It consists of the major divisions of the region with the attribute of possible land-use types in each division. Those major divisions are the spatial objects that provide the boundaries to select only the relevant classes of land-use types for the classification. SORSICT is tested for upcountry tea plantation of Sri Lanka. Quick Bird-2008 Satellite imagery of a selected district of Sri Lanka was used to evaluate the proposed approach. Comparison of the results of SORSICT with the conventional methods reveals higher accuracy in SORSICT at identifying tea plantation from remotely sensed data. SORSICT provides a solution for the problem of generating irrelevant classes in remotely sensed imagery classification by applying a spatial threshold scheme.