The basic assumption for image classification is a specific part of the feature space corresponding to a specific class. Classes have to be distinguished in an image and classification needs to have different spectral characteristics. This can be analyzed by comparing spectral reflectance curves. Image classification gives results to certain level of reliability. The principle of image classification is that a pixel is assigned to a class based on its feature vector by comparing it to predefined clusters in the feature space. Doing so for all image pixels result in a classified image (Janssen. 2001). There are two approaches for classification. One is the pixel based image analysis approach and the other object oriented image analysis approach. For this study Pixel Based Image analysis approach was used. A. Unsupervised Classification: The unsupervised classification was carried out for all the three images. The spectral classes obtained from the unsupervised classification are based solely on natural groupings in the image values. The spectral classes obtained from all the three images were not initially known. So taking the reference values, the classified data was compared and the spectral classes were identified. B. Supervised Classification: Here the image analyst supervises the pixel categorization process by specifying, to the computer algorithm, numerical descriptors of various landcover types present in the image. Training samples that describes the typical spectral pattern of land cover classes are defined. Pixels in the image are compared numerically to the training samples and are labeled to landcover classes that have similar characteristics. (a). Ground Truth and Crop Classification: Ground truth data were collected by integrated use of global positioning system (GPS), geocoded FCC (False Colour Composite) of LANDSAT (December, 2001), LISS III (February, 2004) and ASTER (June, 2004). The topographic maps could not be used as the area comes under the purview of the restricted area. Ground truth information corresponding to various landuse/landcover classes were collected after critically examining the spectral variations in the geocoded FCC. The field trips were carried out from 6th August to 16th August, 2005. The complete enumeration process was followed to obtain the information on the landuse/landcover class. ERDAS Imagine 8.7 was extensively used for image processing and interpretation of satellite imageries in this study. All the images were georeferenced to Polyconic projection using the ground control points and resampled to the required pixel size using nearest neighbour method. The images were mosaiced to cover the entire study area. The image correspond to the study area was then clipped out using the district boundary. The samples for different landuse/landcover classes were defined interactively on the three different images of LANDSAT, LISS III and ASTER based on the homogeneity of the samples like the uniform colour of the images and the information collected during field visits. For crop identification, supervised classification was performed for all the images. All the three classification techniques like the maximum likelihood classification (MLC), parallelepiped and minimum distance to mean classification have been applied for the images and the best classification technique was then found out. It was observed that Maximum Likelihood Classification (MLC) was giving good results as compared to the other two techniques. On the date of acquisition, it was observed that all the tea plants exhibited a range of ground cover, canopy density and affected tea patches. Further using these classified maps the diseased patches were delineated and the shift in the diseases was observed. C. Accuracy Assessment In accuracy assessment the main assumption is that the reference data or field data are correct. Classification accuracy will be determined by using three complementary measures which are based on error matrices or confusion matrix derived from independent field data. The two methods used for accuracy assessment are: a. The Error Matrix: Error matrix is a square with the same number of information classes that will be assessed as the row and column. Numbers in rows are the classification result and numbers in columns are reference data. In the error matrix the elements in the diagonal are pixels that are correctly classified. Error matrix is the most effective way to represent map accuracy. Overall accuracy: It is the proportion of all reference pixels, which are classified correctly. It is computed by dividing the total number of correctly classified pixels by the total number of reference pixels. Overall accuracy can be given by:  Overall accuracy is a coarse measurement. It gives no information about what classes are classified with good accuracy.
 Producer’s accuracy tells how well the classification agrees with reference classification.  User’s accuracy predicts the probability that a pixel classified as Class I is actually belonging to Class I. It is a discrete multivariate technique used in accuracy assessment for statistically determining if one error matrix is significantly different than another (Bishop et.al, 1975). The result of performing a Kappa analysis is a KHAT statistic (k^ , an estimate of Kappa) which is another measure of agreement or accuracy. This measure of agreement is based on the difference between the actual agreements in the error matrix and the chance agreement which is indicated by rows and column totals. (b). Assessing the Accuracy of the Classified Images In this study, accuracy assessment of MLC based crop/landuse classes was achieved in the form of error matrix by comparing classified output with the ground truth information of independent sites, collected using GPS. A total of 375 sample sites/field was geo-located with GPS for comparison with classified landcover types: healthy tea patches, moderately healthy patches, diseased tea patches, scrubs, river, river bed, barren land and settlements. Overall accuracy was defined as the percentage of total independent reference pixels that were correctly classified by the MLC. Producer’s accuracy was calculated by dividing the number of pixels correctly classified for each class by the total number of independent reference pixels for that class while the user’s accuracy was the number of correctly classified pixels divided by the total number of classified pixels for that class. The accuracy results are discussed in the results and discussion chapter. |