|
Object-oriented Approach for Rubber and oil palm Identification
Sutat Dansagoonpon Rubber Research Institute of Thailand Department of Agriculture Email: st027123@ait.ac.th Nitin K Tripathi*, and Roberto S. Clemente** * Remote Sensing & GIS program, Asian Institute of Technology (AIT) ** WEM program, Asian Institute of Technology (AIT) Email: nitinkt@ait.ac.th, clemente@ait.ac.th ABSTRACT: The accuracy of image classification is not encouraging. The problem is mainly due to their complex nature, depending on the cropping system, plant clones-types and ages, time of day, light conditions and weather etc. Most of the algorithms for vegetation classification in the image processing domain are solely based on the use of spectral information They do not consider the geometric characteristics given by the spatial distribution such as to roads network in Oil Palm plantation, bare land-like due to Rubber replanting or Oil Palm new planting and shape etc. which are represented on the image. The human visual image interpretation criteria could be used for classification and be broadly defined by the shape, tone or color, size, pattern, texture, shadow, spatial relationships (association) of the ground targets etc. In this study the object-oriented approach using segmentation together with fuzzy classification, interpreter's knowledge-based, experience, and familiarity with a study area was applied to combine spectral data with ancillary spatial information for rubber and oil palm classification. It is found to be a powerful method to image analysis simulating the efficiency of human eye with the overall accuracy of classification 94.43 % the Kappa Coefficient 0.93. Keywords: Rubber, Oil Palm, object-oriented approach, knowledge-based, fuzzy classification 1. Introduction The accuracy of the traditional images classification has proven inadequate due to the lack of efficient tools to digitally classify the vegetation cover features. The problem is mainly due to their complex nature; land cover can look different, depending on the season, cropping system, plant clones-types and ages, time of day, light conditions and weather etc. Regarding classification methods, supervised methods have to be trained usually either by taking samples or by describing the properties of the classes. Therefore, the class-describing information must be as accurate, representative and complete as possible, which is in most cases effectively impossible. Hence, a class description can only be a general estimation of the desired class properties. Estimating the properties also means assuming a more or less known uncertainty about the class description or a known vagueness about the properties’ measured values. Since they are just a special way of sorting algorithms, their results having to be interpreted this can be tough in some cases and lead to numerous repetitions of the classification with slightly adjusted parameters. Supervised classification method has been used widely in Thailand. This is based only on spectral information and does not take into account geometric characteristics associated with feature class given by the spatial distribution. From previous study using supervised pixel base classification for interpret rubber and oil palm in Southern Thailand it was found that bare land feature; new planting area, replanting area and road are classified as a same class (Figure 1), and rubber replanting area and road net work inside oil palm plantation (Figure 2) etc.  Figure 1 Road and New Planting Area  Figure 2 Rubber Replanting Area and Road Network Inside Oil Palm Plantation Since in human visual image interpretation, the criteria used for classification can be broadly defined by the shape, tone or color, size, pattern, texture, shadow, spatial relationships (association) of the ground targets and road network inside farm. An interpreter's knowledge, experience, and familiarity with a study area also can be contributed to the classification process. The improvement of rubber and oil palm classification accuracy, therefore, much be based on both spectral information and geometric characteristics. To solve this problem an object-oriented classification method utilizing image segmentation and fuzzy classification on the results of segmentation is suggested (Sun et al, 2004). Compared to neural networks, since the classificators used for the class descriptions in object-oriented method are next neighbor and membership functions, the advantage is a transparent and adaptable set of classification rules (Baatz et al 1999). The main requirement of this study is to discriminate between different nature textures focused on natural rubber and oil palm plantation, using both spectral and spatial information. METHODS The object-oriented image analysis software eCognition is used to perform image segmentation, nearest neighborhood classification and the development of semantic rules incorporating object attribute information. As a first step, image objects are extracted by a knowledge-free automatic segmentation of the imagery. Secondly, a fuzzy rule base, using spectral, ancillary data and class related feature, classifies the remaining object. Finally, membership functions are used to model local knowledge for classification refinement. The object-oriented approach is in principal independent of the specific segmentation and classification techniques. The basic processing units are image objects or segments, and not single pixels. Even the classification acts on image objects. One motivation for the object-oriented approach is the fact that the expected result of many image analysis tasks is the extraction of real world complex nature, proper in shape and proper in classification. This expectation cannot be fulfilled by common, pixel-based approaches. Classic soft classifiers, fuzzy classification system that was applied for this study, it is a simple technique which basically translates feature values of arbitrary range into fuzzy values between 0 and 1, indicating the degree of membership to a specific class, where 1 express full membership/probability and 0 express absolutely non-membership/ improbability. A fuzzy rule base allows the formulation of knowledge and concepts in a very efficient way. This way, our expert knowledge or ancillary data about general class descriptions as well as relations of classes can be incorporated into the system. Membership functions used in this study is smaller than form. They offer a transparent relationship between feature values and the degree of membership to a class. Result and Discussion The procedure firstly the whole satellite image of Landsat 5 TM has segmented. Based on this segmented image, then object are evaluated by means of next neighbor and membership functions (Figure 5). In this study mature rubber (>6 years old) and mature oil palm ( >4 years old) are classified by supervised classification and maximum likelihood. Soft classifier, fuzzy classification system, was applied for the analyze image objects of young rubber (2-6 years old) and young oil palm (2-4 years old), bare land tend to be rubber (< 1 year old) and tend to be oil palm (< 1 year old).  Figure 3 Object oriented approach 1. Supervise segment classification (Knowledge free) This step based only on spectral information and does not take into account geometric characteristics associated with feature class given by the spatial distribution. It was found mature rubber (>6 years old) and mature oil palm (>4 years old) are well identified meanwhile the other remaining class are more complicate and intermixing between bare land feature; new planting area and replanting area, young rubber and young oil palm. Hence an interpreter's knowledge, experience, and ancillary data need to be applied. 2. Applied Knowledge base for classification Usually a land use class are more complicated, classes often consist of combinations of conditions connected by operators like “and,” “or” and “not.” At this point, we can integrate all available knowledge about the relations between features and class assignment. The flexible fuzzy approach allows integration of knowledge close to human thinking. This step the knowledge base was applied into fuzzy set to identified 2.1 Bare land tend to be rubber as following. 1 st fuzzy set “high relative border length to mature rubber” HL, and smaller than (Boolean) form 2 nd fuzzy set “not relative border length to mature oil palm” NL, and smaller than (Boolean) form 2.2 Bare land tend to be oil palm as following. 1 st fuzzy set “High relative border length to mature oil palm”, HL is defined by a membership function for the object feature smaller than (Boolean) form. 2 nd fuzzy set “Close distance to mature oil palm border”(a fuzzy set CD is defined). Based on an object-oriented approach, mature oil palm is addressed to be the samples. The distance computes as follows:  2.3 Young rubber (2-6 years old) and young oil palm (2-4 years old) There are three fuzzy set were applied for identification as following.. (1) 1 st Fuzzy set “not relative border length to mature oil palm” NL are defined by a membership function for the object feature smaller than (Boolean) form. (2) 2 nd Fuzzy set “relative border length to mature rubber” HL are defined by a membership function for the object feature larger than (Boolean) form. (3) 3 rd Fuzzy set “relative border length to mature oil palm” HL are defined by a membership function for the object feature smaller than (Boolean) form. The structure of the rule base shows very clearly that the classes land tend to be rubber and land tend to be oil palm, young rubber (2-6 years old) and young oil palm (2-4 years old) are similar. Using hierarchy makes this relation obvious and simplifies definitions. Both contribute to a consistent overall rule base (Figure 4).  Figure 4 The Classified Image using object oriented approach 3. Accuracy Assessment of image classification The classification accuracy assessment was done by comparing the classified images (Figure 10) and their ground truth images. Table 1 shows the producer accuracy comparison indicated that in mature rubber, mature oil palm, young rubber , young oil palm , land trend to be oil palm and land trend to be rubber classes the agreement were 98.85, 99.7, 89.58, 98.66, 83.28 and 82.37 % respectively. The user accuracy comparison of mature rubber, mature oil palm, young rubber , young oil palm , land trend to be oil palm and land trend to be rubber classes, pixel-labeled by the classifier, were in agreement in 87.77, 100, 100, 95.06, 99.46 and 94.24 % of cases respectively. The overall accuracy was 94.43 %. The Kappa Coefficient was 0.93, which means that the image had an accuracy that was 93 % better than would be expected through the chance assignment of pixels to categories. 
Conclusion The object oriented approach using knowledge based segmentation together with fuzzy classification is found to be a powerful method to image analysis as a human eye. However accuracy of classification result can be differentiated and improved based only on interpretator’s experience and knowledge base. The classification problem is mainly due to the complex nature of the study area, land cover can look different, depending on crop clones and ages, land use planning or policy of the study etc. For successful classification using fuzzy rule base a deliberate choice of membership function is crucial. This need interpretator’s knowledge introduce into the system. In this study the knowledge base such as; relations between image object to neighbor sample objects, distance and border, farm shape and size, texture etc are also applied. The better the knowledge about the real system is modeled by the membership function, the better the final classification result. REFERENCES
Page 1 of 1
|