A Computer Aided System For Tropical Leaf Medicinal Plant Identification

The objective of this paper is to develop a computer aided system for leaf medicinal plant identification using Probabilistic Neural Network. In Indonesia only 20-22% of medicinal plants have been cultivated. Generally, identification process of medicinal plants has been done manually by a herbarium taxonomist using guidebook of taxonomy/dendrology. This system is designed to help taxonomist to identify leaf medicinal plant automatically using a computer-aided system. This system uses three features of leaf to identify the medicinal plant, i.e., morphology, shape, and texture. Leaf is used in this system for identification because easily to find. To classify medicinal plant we used Probabilistic Neural Network. The features will be combined using Product Decision Rule (PDR). The system was tested on 30 species medicinal plant from Garden of Biopharmaca Research Center and Greenhouse Center of Ex-situ Conservation of Medicinal Indonesian Tropical Forest Plants, Faculty of Forestry, Bogor Agriculture University, Indonesia. Experiment results showed that the accuracy of medicinal plant identification using combination of leaf features increase until 74,67%. The comparative analysis of leaf features has been performed statistically. It showed that shape is a dominant features for plant identification. This system is very promising to help people identify medicinal plant automatically and for conservation and utilization of medicinal plants. Keywordss: leaf identification, leaf morphology, leaf shape, product decision rule, medicinal plants, probabilistic neural network.


I. INTRODUCTION
Indonesia is a country with a mega biodiversity. For plants diversity, Indonesia has more than 38,000 plants species [1]. In 2001 the Laboratory of Plant Conservation, Faculty of Forestry Bogor Agricultural University (IPB) has recorded 2,039 tropical medicinal plant species from Indonesia forest ecosystems [2]. Unfortunately only 20-22% of medicinal plants are cultivated by people. One of the conservation and utilization of medicinal plants using the technology is developing a medicinal plants identification system automatically [3].
The plant identification is not easy and required an appropriate background with significant experience. Also number of researcher who can identify medicinal plants is very limited. The researcher must bring a book dictionary of plant during plant identification in the field. This caused by utilization of medicinal plants by community is very low. Therefore, we need a computer-based automatically system as a tool to help people identify these various types of the medicinal plants.
There were some researches on plant identification.
Fourier descriptor was used to extract leaf shape of Dicotyledonous leaf class [4]. The accuracy of system is 31.75%. Basic characteristic of leaf morphological and its derivatives have been analyzed by [5]. The accuracy of system is 27.22%. The Local Binary Pattern (LBP) is used to extract texture of house plants [6]. The accuracy of system is 73.33%.
In this research, we develop a computer aided system for Indonesian leaf medicinal plant identification using three features, i.e., morphological, texture, and shape and also use Probabilistic Neural Network (PNN) for leaf identification.

A. Morphological Feature Extraction
Morphological feature consists of two features, basic and derivate. The basic feature used in this research was diameter, area and perimeter/leaf circumference. Three basic features can be combined to get eight derivate features like smooth factor, shape factor, ratio perimeter and diameter, also five features of leaf vein [4].

B.
Texture feature extraction Local Binary pattern was originally designed for texture description. The operator assigns a label to every pixel of an image using the 3x3-neighbourhood of each pixel with the center pixel value and considering the results as binary number. The histogram of the label is used as texture descriptor. Fig. 1 show illustration of the basic LBP operator. To be able to deal with textures at different scales, the LBP operator was later extended to use different sizes.
Defining the local neighborhood as a set of sampling points evenly spaced on cirle centered at the pixel to be labelled allows any radius and number of sampling points. In this research the notation (P,R) will be used for pixel neiborhoods which means P sampling points on a circle of radius R. Fig. 2 show an example of circular neighborhoods. Generally, texture can be characterized by a spatial structure (e.g. a pattern such as LBP) and the contrast (e.g. shaped like a bell that scaling nonlinear variable [8]. PNN is composed of only three layers: the input layer, the pattern layer and the summation layers. The main advantage of using the architecture of PNN training data is easy and very fast.
Weight are not "trained" but assigned. Existing weight will never be alternated but only new vectors are inserted into weight matrices when training. So it can be used in real time.
In this research classifier combination is used to increase

Morphological Feature Classification
The experimental result showed that average accuracy of leaf morphological features only has the average accuracy is 20%.
Some species get low accuracy. Fig 3 shows accuracy of leaf identification for each class using morphological feature.    The experimental result show that average accuracy of plant identification using shape feature is 64%. In this experiment we used Fourier Descriptor. Fig. 7 showed the average accuracy for each class using shape features. The number of species which can be idetified using shape feature more than species using texture or morphological texture.  A.

Feature Combination
The experimental result show that average accuracy of plant identification using feature combination increased up to 74.67% as shown in Fig. 9. The feature combination has better accuracy compare to single fature. In this experiment we used feature combination of morphology, texture and shape. In this research we used product decision rule (PDR) to combine these features. Fig. 11 showed the average accuracy for each class using feature combination.  show that shape feature is most dominant than morphology and texture. This system is very promising to help people identify medicinal plant automatically. Also this system can be used for conservation and utilization of medicinal plants.