This repository is a 1.5 year long university project. Its main goal is to implement a Generative Adversarial Network to create heightmaps to be used in Unreal Engine. Unreal Engine code will generate a terrain from the provided heightmap with the use of common tools used for procedural terrain generation.
GAN will be trained on data provided by NASA Visible Earth.
https://github.com/mgraves236/GAN-heightmap-generation/blob/master/map.png
Files crop.py and remove.py are responsible for data preprocessing. The first one creates a grid made by 256x256 pixels elements and crops the world map image according to it. Elements that are almost uniform in color are redundant to the training process, thus remove.py removes them.
From 14,112 cropped images, only 5,088 are utilized in the training process.
The model is implemented in train.py. It consists of two simple sequential neural networks which act as a discriminator and a generator.
The generator takes a noise vector as an input and modifies the data to achieve the results most similar to the original data. Its architecture is presented below:
The discriminator's task is to recognize the real examples and the generated ones. It was constructed as a simple 6-layers neural network:
The networks were trained for 35,000 epochs with a batch size of 256 with Adam.
The output images from the generator required further processing by applying shape blur with a circular kernel. Examples of generated images are shown below:
Generated heightmaps were loaded into Unreal Engine. An automaterial was implemented using blueprints to automatically color the loaded terrain:
The final result of the project is the colored terrain generated by AI:
