# 多媒體作業
Generating Terrain Images: Crafting Virtual Training Grounds for Unmanned Drone Flight Training

## introduction
In recent years, drones have been widely used in our daily lives. Their dexterous appearance enables them to shuttle between places easily. Drones can be used not only for taking photos but also for rescue operations. However, before that, we need to train our drones. Training them in natural settings may cause a lot of wear and tear on the machines. Additionally, some rugged terrain is difficult to find. This will cost us a lot of time and money. Therefore, we want to generate different terrain pictures to train drones. If our plan works, we can save a lot of costs from training.

In response to these challenges, this research explores the innovative approach of utilizing Generative Adversarial Networks (GANs) to generate realistic terrain images. By crafting virtual training grounds that closely mimic diverse real-world landscapes, this initiative aims to revolutionize unmanned drone flight training. These virtual environments offer a controlled yet authentic space for pilots to develop and hone their skills in a variety of terrains, scenarios, and weather conditions.

## System framework
The proposal’s objective is to build a system to generate terrain images for unmanned drone flight training. GANs are prevalent model to generate new, realistic data in a vast dataset. GANs consist of two neural networks, generator and discriminator. Generator can generates new data instances. It takes random noise as input and transforms it into data that ideally is similar to the real data. The other network Discriminator evaluates the generated data and real data, distinguishing between them. Its goal is to correctly identify whether a given sample is real or generated.


Since we are going to generate terrain images, we will focus on generating height maps in gradient noise way. Gradient noise is a type of structured, coherent noise used in computer graphics, procedural texture generation, and various other applications. It is characterized by smoothly varying patterns and is often employed to create natural and realistic textures. Moreover, Gradient noise is widely used in procedural texture generation, where it contributes to the creation of intricate and realistic patterns without the need for explicit texture maps.

## Expected results
Our model is expected to generate high-quality terrain images that closely resemble real-world landscapes. Furthermore, the generated terrains should exhibit diversity to simulate different geographical features, allowing for comprehensive drone flight training scenarios.

To achieve this goal, we will collect data as more as possible. We plan to use NASA’s SRTM for our extra dataset. NASA’s SRTM is the maps that consist of the majorityoftheplanet’s elevation data. A geo-rectangle of longitude-latitude values were used to define regions that would then gather and populate a height grid with the values for the corresponding points from the SRTM data set.

Generally, our ultimate target of this research endeavor aim to contribute to the development of advanced drone training methodologies by leveraging state-of-the-art technology in terrain image generation. The outcomes should demonstrate the potential for cost-effective, realistic, and adaptable virtual training environments for unmanned drone flight training.

## 參考文獻
[Automatic path generation for terrain navigation](https://www.sciencedirect.com/science/article/pii/S0097849312001422)
[Realistic and Textured Terrain Generation using GANs](https://eprints.whiterose.ac.uk/153088/1/Real_world_Textured_terrain_generation_using_GANs_1_.pdf)
[Deep Convolutional Generative Adversarial Network for Procedural 3D Landscape Generation Based on DEM](https://link.springer.com/chapter/10.1007/978-3-319-76908-0_9)
[Procedural Terrain Generation Using Generative Adversarial Networks](https://www.researchgate.net/publication/356886368_Procedural_Terrain_Generation_Using_Generative_Adversarial_Networks)

[pix2pix-terrain-generator](https://github.com/jayin92/pix2pix-terrain-generator)