Improvement in recognition accuracy of minority crops by resampling of imbalanced training datasets of remote sensing

Abstract

The rapid development of high-spatial-resolution satellites has effectively alleviated the problem of mixed pixels in satellite images, thereby enabling extraction of the meticulous distribution of crops from them. The classification of remote sensing images is a quick way to obtain accurate agricultural information. However, the accuracy of supervised classification using remote sensing images is affected by several factors, such as classifier algorithm and input datasets. The imbalanced training samples, which indicates the number of training samples of some categories is considerably smaller or larger than the others, often results in poor classification accuracy for the minority classes. To improve this situation and generalization performance of classifier, this research focused on proper utilization of resampling techniques and classification methodologies for achieving perfect performance of remote sensing image classification. We investigated the aforementioned images by data mining approaches including spectrum and texture features and selection of optimized features based on recursive feature elimination. Then, five resample methods, namely, three over-resampling methods and two under-sampling methods, were separately used to balance the initial training datasets. Finally, we tested the resampled datasets by utilizing two classifiers (decision tree and AdaBoost) and evaluated the performance of each one in terms of kappa coefficient, overall accuracy, producer's accuracy, and user's accuracy. The overall classification accuracy and kappa coefficient improved considerably on decision tree (14.32%) and AdaBoost classifier (10.23%) after resampling. The AdaBoost obtained the highest value of kappa coefficient (0.9336) by using the training dataset resampled with ADASYN. The accuracy of classification on minority crops was also increased by resampling training datasets. Meanwhile, feature selection results showed that vegetation and texture indexes were more efficient than features of original reflection ratio to classification. Over-resampling methods had advantages in relieving the influence of imbalanced training samples to classifiers. Resampling process to training datasets has remarkable advantage in improving the classifier performance if the training datasets are critically imbalanced. The detailed accuracy assessment shows that over-resampling method is more excellent than under-resampling. The reason is that some significant samples are lost during under-resampling, but helpful and useful information is added after over-resampling. AdaBoost classifier performs better than decision tree in terms of solving imbalanced training datasets. Combination of proper resampling approaches and compatible classifier can significantly improve the accuracy of minority classes in the situation of imbalanced dataset classification.