As we said in the last article Discussion Chapter, our project is still in its first stage of development and could be improved in several ways. First of all, we have to refine our research question. We wanted to know if the model was capable of generating building’s footprints for cities with less construction density. The population density wasn’t a good parameter to measure the building’s density since it couldn’t represent the building’s distribution accurately. Therefore we could base our research on other indicators of building’s density such as the amount of construction per tiles and focus only on the area with building footprints after having created the XYZ tiles repartition of the map, but the best indicator for our model to learn the urban morphology remains the street network. In fact, a lot of future work would be to study the street network to choose the best types of street for our model and exclude the one that could confuse it. We could also add colors to other street types to see if more colors could help the model to differentiate street’s types.
Once again, as the authors of GANmapper did, we could also train our model with different zoom sizes to find the best one. In addition, it would also be interesting to change some of the hyperparameters such as the metrics for evaluation. In the article “Pros and cons of Gan Evaluation measures : New Developments”, tha author Ali Borji shows a large list of multiple and various metrics for evaluation. In our study case, we could replace the normal FID with a fast FID to improve the learning rate or a clean FID to make it more accurate. At last, we have to train our model way more to make it learn and see better results. Another important part of our future research would be to study the training behavior of conditional GAN. With a better understanding of the system, it will be easier to adapt certain parameters to improve our model.
For future projects we could use our model on villages and small cities from the countryside and check if the model is able to understand a less dense urban morphology. We could also try to generate points of interest as the authors of GANmapper recommended in their article. As a result, there’s still a lot of work to be done and we’ll keep you informed of our progress, so stay tuned!