Integrating multiple texture methods with random forest classifiication algorithm to classify spectrally smillar agricultural crops

Hazelnut and tea are two main crop types cultivated in Trabzon, Sürmene. This dissertation aims to determine extent and distribution of hazelnut and tea grown areas, fast and accurately using satellite images taken over Trabzon, Sürmene region. In this study, Worldview-2 satellite images , which have eight multi spectral bands (MS:2m) and one high spatial resolution panchromatic band (PAN:0.5m) were used. Since the study area contains agricultural products, which are grown in different seasons, satellite images belonging both summer and winter periods were used. These images were classified with machine learning based algorithms such as Random Forest (RF), Support Vector Machine (SVM) and Gentle AdaBoost (GAB) as well as classical statistical based Maximum Likelihood Classification (MLC) method, and the results were evaluated. For summer period, preliminary results acquired using only spectral values indicated that RF with 79.05% overall accuracy gives higher classification accuracy than other methods, that is ~7%, ~10% and ~19% better accuracy than GAB, SVM and MLC, respectively. For winter period, results indicated that RF, with 71.84 % overall accuracy, is also more successful other methods, that is ~7%, ~%6 and ~%8 better accuracy than GAB, SVM and MLC, respectively. Furthermore, integration of different feature extraction methods such as co-occurrence matrix, gabor, curvelet, NDVI vegetation index and Digital Elevation Model (DEM) and their contributions to success of RF classification method was examined. Gabor filter and NDVI index, which are the most successful ones among these methods, improved the overall accuracies of the RF around 9% for summer period and around 8% for winter period. Then, thematic crop maps and a data base were generated using classified images. Finally, produced thematic maps were compared with up to date and cadastral maps to validate classification results with ground truth data.