A Dual Phase Approach for Addressing Class Imbalance in Land-Use and Land-Cover Mapping From Remotely Sensed Images

Semantic segmentation of remotely sensed images for land-use and land-cover classes plays a significant role in various ecosystem management applications. State-of-the-art results in assigning land-use and land-cover classes are primarily achieved using fully convolutional encoder-decoder architectures. However, the uneven distribution of the land-use and land-cover classes becomes a major hurdle leading to performance skewness towards majority classes over minority classes. This paper proposes a novel dual-phase training, with the first phase proposing a new undersampling technique using minority class focused class normalization and the second phase that uses this learnt knowledge for ensembling to prevent overfitting and compensate for the loss of information due to undersampling. The proposed method achieved an overall performance gain of up to 2% in MIoU, Kappa, and F1 Score metrics and up to 3% in class-wise F1-score when compared to the baseline models on Wuhan Dense Labeling, Vaihingen and Potsdam datasets.