Tree species classification from high resolution digital orthophoto maps

Advancement in remote sensing (RS) technology has made it possible to acquire images with high spectral and spatial resolutions. It offers an alternate solution to the traditional approaches of field surveys and manual photo interpretation, which are both costly and labor intensive. On the other hand, images acquired using multi/hyperspectral sensors usually provide enough detail information so that objects even at fine scale can be obtained, which can be used in various applications, such as management and monitoring of forest and agricultural resources. However, the amount of data acquired is so large that manual interpretations by human experts become impractical in many situations. Especially, the task of tree identification is challenging due to complexity of vegetation information and spectral similarity among various species. RS data has brought both opportunities and challenges which requires development of new processing techniques to effectively extract tangible information. The focus of this research is to investigate an object-based approach based on mean shift and combination of supervised classifiers for detection of tea gardens from high resolution digital orthophoto maps obtained from an airborne sensor. The proposed method was carried out in several sequential steps. In the first step, the semantic relationship between pixels was exploited and transformed the raw pixels into an object-based representation by using the spatial relationship among pixels. Mean shift based clustering algorithm, which is a non-parametric mode seeking algorithm, was used for delineation of image objects. These objects were then used as building block for the rest of processing. Image filtering was then applied to remove unwanted noisy elements from the object-based representation before extracting features from individual objects. Morphological processing was performed to get smooth object borders and possibly remove small holes within the objects. Features were derived from spectral, spatial and texture domains from each object and then normalized to train and test the classifiers. Several features were considered, however, only those features were selected which are discriminative and optimized the classification results. Sequential forward selection method was used with Jeffries-Mautasia (JM) distance metric to select highly discriminant feature set. Three most widely used supervised classifiers were selected for this study, namely support vector machine (SVM), artificial neural network (ANN) and random forest (RF). These classifiers belong to a diverse family of statistical learning and they are proven to be effective for large scale data classification. To train the classifiers, training sample were selected for tea gardens and other types of trees. A brute force grid search algorithm was applied to find optimal parameters for each classifier using training and validation set. The outputs obtained from each individual classifier was combined using a maximum voting approach to produce the final output as thematic map for tea gardens. Finally, experiments were performed to evaluate the effectiveness of the proposed method for classification of tea gardens from high resolution digital orthophoto maps by comparing with manually digitized tea gardens images. It was shown that integration of object-based segmentation with multi-classifier approach is an effective method for extraction of tea gardens with high classification accuracy.