研究目的

To address the problems arising from random sampling strategies in conjunction with spectral-spatial and pixel-wise classifiers such as 3D Convolutional Neural Networks (3D CNN) and propose improved sampling strategies based on the Density-Based Clustering Algorithm (DBSCAN) to minimize the violation of the train and test samples independence assumption.

研究成果

The study confirms that random sampling approaches violate the independence assumption due to the introduction of systematic bias, particularly for current state-of-the-art CNNs and the spatial overlaps in their receptive fields. The proposed sampling approaches using the DBSCAN clustering algorithm minimize said bias and result in a classification accuracy on unseen test data closer to an actual out-of-sample performance. The findings suggest a more wide-spread adaptation of non-random sampling approaches for remote sensing classification problems.

研究不足

The study is limited to the Indian Pines dataset and the proposed sampling strategies may introduce bias when employing the metric on all clustered regions before splitting them into training and test data. The generalization of the findings to other datasets and the influence of the proposed sorting metrics for the regions on the classifier performance require further investigation.