AUTOMATIC TRAINING SITE SELECTION FOR AGRICULTURAL CROP CLASSIFICATION: A CASE STUDY ON KARACABEY PLAIN, TURKEY

Abstract. This study implements a traditional supervised classification method to an optical image composed of agricultural crops by means of a unique way, selecting the training samples automatically. Panchromatic (1m) and multispectral (4m) Kompsat-2 images (July 2008) of Karacabey Plain (~100km²), located in Marmara region, are used to evaluate the proposed approach. Due to the characteristic of rich, loamy soils combined with reasonable weather conditions, the Karacabey Plain is one of the most valuable agricultural regions of Turkey. Analyses start with applying an image fusion algorithm on the panchromatic and multispectral image. As a result of this process, 1m spatial resolution colour image is produced. In the next step, the four-band fused (1m) image and multispectral (4m) image are orthorectified. Next, the fused image (1m) is segmented using a popular segmentation method, Mean- Shift. The Mean-Shift is originally a method based on kernel density estimation and it shifts each pixel to the mode of clusters. In the segmentation procedure, three parameters must be defined: (i) spatial domain (h_s), (ii) range domain (h_r), and (iii) minimum region (MR). In this study, in total, 176 parameter combinations (h_s, h_r, and MR) are tested on a small part of the area (~10km²) to find an optimum segmentation result, and a final parameter combination (h_s=18, h_r=20, and MR=1000) is determined after evaluating multiple goodness measures. The final segmentation output is then utilized to the classification framework. The classification operation is applied on the four-band multispectral image (4m) to minimize the mixed pixel effect. Before the image classification, each segment is overlaid with the bands of the image fused, and several descriptive statistics of each segment are computed for each band. To select the potential homogeneous regions that are eligible for the selection of training samples, a user-defined threshold is applied. After finding those potential regions, the training pixels are automatically selected and labelled. Thereafter, those training pixels are utilized in a traditional Maximum Likelihood Classification to classify five crop types namely; corn, tomato/pepper, rice, sugar beet, and wheat. The accuracy of the classification is evaluated in pixel-based manner with the help of a reference map including crop information of the area. Promising results are achieved for pixel-based approach. Based on the error matrices used in the evaluation, overall accuracy of the pixel-based analysis is computed as 89.31%. Similar to the overall accuracies, high individual class accuracies are obtained as well. The results point out that automatically collecting the training samples by extracting representative homogenous areas significantly increases the speed of the classification and minimizes the human interaction. The results also confirm that the proposed approach is highly appropriate for the extraction of representative homogenous training areas.