Weed classification using genetic algorithm optimised classifiers
Automated spot weeding with an efficient weed classification can increase production in crops and reduce herbicide usage. A proposed strategy of applying excessive feature sets followed by feature selection was applied on development of the classifiers to eliminate the non-discriminating features...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2016
|
| Online Access: | https://eprints.ums.edu.my/id/eprint/12163/1/Weed%20classification%20using%20genetic.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Automated spot weeding with an efficient weed classification can increase
production in crops and reduce herbicide usage. A proposed strategy of applying
excessive feature sets followed by feature selection was applied on development of
the classifiers to eliminate the non-discriminating features. Artificial Neural Network
(ANN) and Support Vector Machines (SVM) were applied in the classification using
a combination of image derived features. Optimising the classifier involves a
tedious selection of subsets and parameters which can be considered a solution
searching problem. Optimising the classifier parameters can be solved using
heuristic methods such as Genetic Algorithm since it is a non - convex optimisation
problem. In ANN structures, the features subset (input numbers) and hidden
neuron layer are configurable while for SVM, the hyper parameter and the feature
subset are configurable. In order to optimise the structures, feature subset and
parameters, two optimisation approach were considered. These two optimisation
approach include using backward Sequential Feature Selection (SFS) and Genetic
Algorithm (GA) approach. GA requires a careful design of chromosome and fitness
function in representing the structure, parameters and feature sets. In the fitness
function for SVM optimisation, the fitness score is weighted between feature
reduction term and fitness evaluation term of the candidate solution. For the SVMs
optimised with GA, it was observed that all the GA configurations yielded better
results (both on validation/test sets) as compared to SFS optimised counterpart.
The results suggest that optimisation fitness function for SVM requires a
simultaneous selection of feature subset /hyper parameters and a small value of
weightage (between 0% to 20%) of the total fitness score should be allocated
from the feature reduction term to avoid over fitting to training sets. As for the
ANN optimisation using GA, fitness function (which includes the error reduction
term, feature reduction term and neuron reduction term) showed lesser
generalization with independent test sets in comparison with the SFS optimisation
approach. The ANN configuration with SFS feature selection gave best results on
validation error therefore showing better subset selection using SFS algorithm as
compared to GA selection. |
|---|