Innovation Lab Use Case

Optimize Choice in Neural Network

Several state-of-the-art convolutional neural networks were modified for the segmentation of very high-resolution satellite imagery. The applicability of these networks and the effects of underlying hyperparameters was evaluated in the classification of features in 0.5-meter resolution data. The findings were that:

• Some of the most common networks (shown below) are not designed to identify small and sparse objects.
• Training and inference time increases drastically as the network integrates more blocks.
• For Residual Neural Network (ResNet), the bigger the network, the more accurate it becomes.
• Atrous convolutions, as they are, were found to be inefficient at detecting individual pixels that were outside of a small class group.