Prediction Of Leaf Mechanical Properties Based On Geometry Features With Data Mining

• Main Author: H’ng, Choo Wooi
• Format: Thesis
• Language: English
• Published: 2019
• Subjects: T Technology; TJ1-1570 Mechanical engineering and machinery
• Online Access: http://eprints.usm.my/47859/1/Prediction%20Of%20Leaf%20Mechanical%20Properties%20Based%20On%20Geometry%20Features%20With%20Data%20Mining.pdf; http://eprints.usm.my/47859/

The leaf mechanical properties are typically determined by mechanical tests to study the leaf‟s lifespan, its anti-herbivore defences and the ecological functions. The influences of habitats, environmental resources such as nutrient, light, and water, and species diversity on the leaf anatomies and their chemical compositions were previously considered. However, the mechanical properties of the leaves from the geometry and morphology aspects are still vague. The main goal of this study is to examine the effect of various geometrical attributes to predict the leaf mechanical properties based on four different indicators using data mining approach. An experimental study involving 20 different species of the terrestrial plants were conducted. A total of 600 x 23 features attributes comprising of leaf geometrical features, discriminant features and its derived quantities were collected by measurements, field observations and the tearing test performed using the Universal Testing Machine (UTM). The recorded data were screened on data normalization while the outliers were discarded prior to regression analysis aided by the Waikato Environment for Knowledge Analysis (WEKA) tool. The leaf mechanical property indicators: Tearing Force (FT), Tearing Strength (ST), Work-to tear (WT), and Specific Work-to-tear (SWT) identified were predefined as the numeric class attribute. The leaf mechanical properties indicators were predicted using the GaussianProcess, LinearRegression, MultilayerPerceptron (MLP), SMOreg, M5P and REPTree algorithms of WEKA tool, verified on Root Relative Squared Error (RRSE) evaluation index. Findings showed that the numerical predictions on FT and ST (RRSE ~ 25%) were about two folds better than the WT and SWT (RRSE ~50%) in the six algorithms tested. The best prediction performance was gained on FT indicator using the M5P algorithm (RRSE = 22.44%). The linear models and rules developed from the M5P algorithm were adopted for the FT indicator prediction modelling of 14 attributes. The „Species‟ attribute contributes the most for the M5P regression model. Findings also indicate that leaf mechanical properties were insufficient to be represented by its geometry features alone. The M5P regression model was further simplified into 9 attributes showing insignificant difference determined on the paired T-test between the RRSE achieved by M5P regression and the simplified model (RRSE = 21.37%).