Improve the Recognition Accuracy of Minor Crops By Resampling with Imbalanced Training Data of Remote Sensing
The rapid development of high spatial resolution satellites has effectively alleviated the problem of mixed pixels in remote sensing image data. Which makes it possible to get the meticulous distribution of crops from remote sensing images. The classification of Remote Sensing images is a quick way to obtain accurate agricultural information. However, the accuracy of supervised classification of Remote Sensing images is usually affected by several factors such as the classifier algorithm and the input data sets. The imbalanced training samples dataset, which means the number of training samples of some classes is much smaller or bigger than the others, often results in poor classification accuracies for minor classes. To solve this problem and improve the generalization performance of classifier, this research focuses on the proper utilization of resampling techniques and classification methods to achieve the best performance in Remote Sensing images classification.
In this study, five resample methods, including three over-resampling methods and two under-sampling methods, are used for balancing the original training dataset. And the thesis deals with the Remote Sensing images by data mining including spectrum and texture and selecting the optimized features bases on RFE (Recursive Feature Elimination). Finally, we test the resampled data sets by utilizing two classifiers (The Decision Tree and Adaboost) and evaluate the performance of them through kappa coefficient, commission and omission.
The results showed the overall classification accuracy and kappa coefficient have been improved a lot on both decision tree and Adaboost. Respectively, the Decision Tree: 14.32% The Adaboost: 10.23%. And the Adaboost gets the highest value of kappa coefficient of 0.9336 by using the training dataset resampled with Adaptive synthetic sampling approach (ADASYN). What’s more, the accuracy of classification on minority crops also better than before. Additional, the result of feature selection shows vegetation indexes and texture indexes is more efficient than features of original reflection ratio of bands. Over-resampling has more advantages to relief the influence of imbalanced training samples to classifier on ameliorating ability to hand with of classifier.