Impact of Tile Size and Tile Overlap on the Prediction Performance of Convolutional Neural Networks Trained for Road Classification

Abstract

Popular geo-computer vision works make use of aerial imagery, with sizes ranging from 64 × 64 to 1024 × 1024 pixels without any overlap, although the learning process of deep learning models can be affected by the reduced semantic context or the lack of information near the image boundaries. In this work, the impact of three tile sizes (256 × 256, 512 × 512, and 1024 × 1024 pixels) and two overlap levels (no overlap and 12.5% overlap) on the performance of road classification models was statistically evaluated. For this, two convolutional neural networks used in various tasks of geospatial object extraction were trained (using the same hyperparameters) on a large dataset (containing aerial image data covering 8650 km2 of the Spanish territory that was labelled with binary road information) under twelve different scenarios, with each scenario featuring a different combination of tile size and overlap. To assess their generalisation capacity, the performance of all resulting models was evaluated on data from novel areas covering approximately 825 km2. The performance metrics obtained were analysed using appropriate descriptive and inferential statistical techniques to evaluate the impact of distinct levels of the fixed factors (tile size, tile overlap, and neural network architecture) on them. Statistical tests were applied to study the main and interaction effects of the fixed factors on the performance. A significance level of 0.05 was applied to all the null hypothesis tests. The results were highly significant for the main effects (p-values lower than 0.001), while the two-way and three-way interaction effects among them had different levels of significance. The results indicate that the training of road classification models on images with a higher tile size (more semantic context) and a higher amount of tile overlap (additional border context and continuity) significantly impacts their performance. The best model was trained on a dataset featuring tiles with a size of 1024 × 1024 pixels and a 12.5% overlap, and achieved a loss value of 0.0984, an F1 score of 0.8728, and an ROC-AUC score of 0.9766, together with an error rate of 3.5% on the test set.