This paper presents the contribution of the ISAR group at RMIT University to the ImageCLEF 2011 Plant identi cation task. The task involves identifying various di erent species of trees based on images of their leaves. Our main objective is to investigate the performance of two classi cation algorithms in associating the correct tree species to each test image. We extracted visual features from the data set using the feature extraction module in GIFT. From all the features extracted, we selected 166 features of the colour histogram. The classi cation algorithms used are instance based learning and decision trees. Both algorithms were implemented using the Weka 3 data mining toolkit. Classi ers for both algorithms were evaluated by a 10 folds cross-validation. Based on the o cial results, our runs did not perform well due to three main reasons namely, feature selection, training data and classi er parameters.
This paper presents the participation of the ISAR group (Information Storage Analysis and Retrieval) at RMIT University in the ImageCLEF 2011 Plant Identi cation task. The task was motivated by the need to accurately gather knowledge of the identity, geographic distribution and uses of plants in ensuring advancement in agriculture and safeguarding its diversity.
The main goal of the task is to correctly associate tree species to each test image. The task is treated as a supervised classi cation problem with tree species used as class labels. Our objective in the task, however, is to investigate the performance of two classi cation algorithms in classifying the test images to the tree species.
The pilot task dataset contains approximately 5400 pictures of leaves from
71 tree species from French Mediterranean area. Further details regarding the
general setup of the dataset are available in the task description [
2.1
Feature Extraction
Classi cation can be done by either using textual features (from the XML les),
visual features (from the jpg les) or combination of both textual and visual
features. Our work is based on visual features only. We extracted the visual
feature from the data set using the feature extraction module in the GNU
ImageFinding Tool (GIFT) [
The classi ers were trained using all the training data together, without separating them according to the type of image acquisition. In order to reduce variability and estimate how accurately the classi er will perform, they were evaluated by a 10-folds cross validation. Note that Weka support several instance based algorithms namely IB1 and IBk whereby k = 2,....,n. Table 1 shows the classi cation rate for the di erent classi ers trained.
From the table, we can see that decision tree and instance based learning classi er performs better than the rest. Although IBk performed slightly better than J48, we selected IB1 and J48 so as to compare between the two di erent classi ers.
The IB1 algorithm is identical to the nearest neighbours algorithm. It is
considered as a statistical learning algorithm and is simple to implement. When
asked to make a prediction about an unknown point, the nearest-neighbour
classi er nds the closest training-point to the unknown point and predicts the
category of that training-point accordingly to some distance metric [
A decision tree partitions the input space of a data set into mutually exclusive regions, each of which is assigned a label, a value or an action to characterise its data points. It is used to classify a case by starting at the root of the tree and moving through it until a leaf is encountered. At each non-leaf decision node, the case's outcome for the test at the node is determined and attention shifts to the root of the sub-tree corresponding to this outcome. When this process nally leads to a leaf, the class of the case is predicted to be that recorded at the leaf.
The decision tree mechanism is transparent and we can follow a tree structure
easily to see how the decision is made. However, many decision tree construction
algorithms involve a two-step process. First, a very large decision tree is grown.
Then, to reduce its large size and over- tting the data, in the second step, the
given tree is pruned [
The Weka 3 implementation of IB1 classi er uses normalised Euclidean distance to nd the training instance closest to the given test instance, and predicts the same class as this training instance. If multiple instances have the same distance (closest) to the test instance, the rst instance found is used. The di erence between IB1 and IBk is that there are no parameters that could be changed.
As for the decision tree classi er, J48 is Weka's implementation of the C4.5 algorithm. The C4.5 decision tree can be either a pruned or unpruned tree. In our experiment, we created a pruned decision tree. We did not use binary splits when building the trees. The con dence factor used for pruning the tree was 0.25 with the minimum number of instance per leaf set as 2. In determining the amount of data used for pruning (number of folds), we used the default value 3. We considered the subtree raising operation when pruning and did not smooth the object counts at the leaves. 3
The objective of our experiment is to evaluate the e ectiveness of both our classi ers in classifying tree species based on images of its leaves. We submitted two runs, one for each classi er. RMIT run1 used the instance based learning IB1 classi er while RMIT run2 used the decision tree classi er J48. As shown in Table 2, our rst run, RMIT run1 performed slightly better in terms of average images identi ed for each type of image acquisition. However, it was unable to identify images of the scan-like type. Our group submitted two runs in our rst participation in the ImageCLEF 2011 Plant identi cation task. The results obtained by our runs were poor. This is due to three main reasons. The rst is poor selection of features. We only used visual features which is the colour histogram and it was not suitable in identifying images based on the type of image acquisition used in the task. Next, we used all the training data together to train the classi er instead of dividing them according to the type of image acquisition. Finally, we did not exhaust all the parameters used in training both of the classi ers. We hope to further improve our experiment in future tasks.