Introduction:
Nimisha was doing her research on surface water connectivity mapping. She needed to map surface water in the Canadian shield and see the things that cause switch on and switch off in the movement of water. The Canadian shield contains very narrow rivers and lakes with <10m resolution. DEM is of great importance to geographers in performing hydrological modelling and simulation. They currently have access to arctic DEM - a dataset with a high resolution of 2m. However, due to large gaps and jumps in the arctic DEM data, they choose not to use it for research and analysis. Instead, they rely on Merit DEM, which has a low resolution of 90m, which is not sufficient to map the area. Our team’s focus is to increase the resolution of and thereby resolve “choke points” in the lower resolution Merit DEM. Choke points are the points which cause switching off of the water when there is low flow and switching on the water when there is high flow. They are an essential component in understanding the hydrological process. If our project is successful and produces great results, it would have a great impact in the field of hydrology. We are using the CNN method for reconstructing the Super Resolution of DEM.
Challenges:
What has been the hardest part of the project you’ve encountered so far? The hardest part of our work is implementing the architecture in the area where the elevation difference is not more. This is creating fuzziness in our images. Also, the grid line has appeared in the reconstructed DEM. Also finding the right loss function was also a challenge for us. Getting a model that performs well. Results are not satisfactory.
Insights:
We expected to get sharp, high resolution predicted images which are close to the original high-resolution images but the model is not performing as expected. It is getting fuzzy images with grid patterns in the image. We think it’s because we are doing RMSE loss on output image as opposed to on output image features.
Plan:
Refining the model architecture and loss function. We still haven’t figured out the R-CNN implementation described in Mukherjee et al., paper as pictured below. Currently we have a vanilla CNN. Also we need to dedicate more time to adding hill shade and slope bands to DEM and see the result of the data in order to make the model learn the hillshade and slope so that it can learn better and we will be able to get a good result.
Next steps:
What are you thinking of changing, if anything? Mentioned above - loss function and CNN -> RCNN or something of the like.
Log in or sign up for Devpost to join the conversation.