Using Monoscopic Multispectral Earth Observation Images to Predict Terrain Features With Deep Neural Networks

In the field of remote sensing and Earth observation, deep neural networks (DNNs) have established themselves as important tools for many different image analysis applications. Estimation of terrain features from optical satellite imagery is a rarely studied application for which DNNs are well suited because of their ability to extract and combine information at various scales. To predict terrain slopes in optical images, we propose an R2U-Net using global multispectral Sentinel-2 (S2) L2A images as input, and ALOS World 3D DSM elevation data as target data. The R2U-Net takes advantage of a residual unit that benefits deep architecture training, and the recurrent residual convolutional layers provide better feature accumulation. Two models were experimented with; one model trained on only the optical RGB bands and one model trained on all S2 L2A bands. Evaluation of the multispectral- and RGB-trained models showed that the multispectral-trained model performs better than the RGB model, both during training and when evaluated on the test data. The multispectral model performs better overall than the RGB model in all the cases studied. Slope errors typically increase from low- to high-gradient terrain, but not at the same rate as the slope steepness itself, while aspect errors decrease as the models struggle more to predict the slope aspect in low-gradient terrain. This highlights that, in this case, using more spectral bands when predicting terrain slopes helps improve the model predictions. The results have also been shown to depend on the incoming angle of the sunlight, which is mostly due to the topographic shadows that are being cast onto the terrain.