Applying Genetic Algorithm-Based Feature Selection and Bimodal Distribution Removal Technique to Classify Dry Sclerophyll Forest Dataset

The forest classification will be efficient if the neural network can be trained well with the features provided by the satellite and this will probably reduce heavy human-based investigation on finding forest regional distributions and obtain forest distribution in time. In this paper, the basic ANN model has been provided with two different techniques applied for network optimization and then the analysis of reasons why these two accuracies vary between two methods will be explained. The basic artificial neural network for accomplishing the dry sclerophyll forest classification task is implemented with the bimodal distribution removal (BDR) technique to learn the impacts of three variables in the network, namely, inputs, weight factors, and learning rate and then compare results with the genetic algorithm to the same dataset. To analyze the impact of the noisy attributes, the other technique feature selection based on the genetic algorithm is adopted for optimizing the previous model by reducing the computation time and then finding out whether the inner relations can be removed based on the fitness function after fitting the inputs into the model. Results can be considered as evidence to support both the positive impact of the BDR on training the dataset as a similar result can be derived like that given by Gedeon with few computations on inputs, and the optimal effectiveness of the feature extraction on the dataset, especially the GIS dataset which contains four noisy data columns, that is, the outputs of the other four kinds of trees. Compared with previous works on the same task classification, this result not only ensures that the BDR can solve forest classification task with similar but relatively lower testing performance and higher accuracy can be reached once the dataset is enlarged to suit the removal technique, but also provides me a chance to deduct the influence of the noisy input on the binary classification problem when the feature selection implemented to reduce computation efforts. Also, compared with repetitive training, testing results, and the cross-validation after processing the models with different assigned values to variables in the network, enhanced confidence to ensure the network performs well when one column noise has been added and my knowledge of how to evaluate the overfitting degree especially when several features are removed.