Grid-Shift: An Image Preprocessing Approach to Reduce Overfitting in AI Training

We present Grid-Shift, a lightweight image pre-processing approach to counteract overfitting when training Convolutional Neural Networks. Grid-Shift solves the problem that tiling large images for training disrupts coherent features (i.e. an object may be split at the edge of a sub-image) and thus leads to information loss. Existing augmentation methods that reduce overfitting do not solve this problem explicitly. In our case study of Land Use and Land Cover Classification, Grid-Shift outperforms all other approaches tested (a raw UNet, a UNet with Batch Normalization, and various augmentation methods). Grid-Shift achieves a Categorical Accuracy of 95%, which is almost 20% better than a raw UNet and still 4% better than the best augmentation approach tested.