Stratified Sampling-Based Deep Learning Approach to Increase Prediction Accuracy of Unbalanced Dataset

Abstract

Due to the imbalanced nature of datasets, classifying unbalanced data classes and drawing accurate predictions is still a challenging task. Sampling procedures, along with machine learning and deep learning algorithms, are a boon for solving this kind of challenging task. This study’s objective is to use sampling-based machine learning and deep learning approaches to automate the recognition of rotting trees from a forest dataset. Method/Approach: The proposed approach successfully predicted the dead tree in the forest. Seven of the twenty-one features are computed using the wrapper approach. This research work presents a novel method for determining the state of decay of the tree. The process of classifying the tree’s state of decay is connected to the issue of unequal class distribution. When classes to be predicted are uneven, this frequently hides poor performance in minority classes. Using stratified sampling procedures, the required samples for precise categorization are prepared. Stratified sampling approaches are employed to generate the necessary samples for accurate prediction, and the precise samples with computed features are input into a deep learning neural network. Finding: The multi-layer feed-forward classifier produces the greatest results in terms of classification accuracy (91%). Novelty/Improvement: Correct samples are necessary for correct classification in machine learning approaches. In the present study, stratified samples were considered while deciding which samples to use as deep neural network input. It suggests that the proposed algorithm could accurately determine whether the tree has decayed or not.