A Modular Generative Approach for Realistic River Deltas: When L-Systems and cGANs Meet

Luis Oswaldo Valencia-Rosado,Zobeida J Guzman-Zavaleta,Oleg Starostenko

doi:10.1109/access.2022.3140226

Luis Oswaldo Valencia-Rosado, Zobeida J Guzman-Zavaleta + Show 1 more

Open Access

https://doi.org/10.1109/access.2022.3140226

Copy DOI

Abstract

Procedural terrain generation aims to create topographically coherent landscapes with realistic terrain features. Realistic landscapes of our blue planet are not complete without river deltas; however, there is an insufficient advancement in the generation of landscapes with this terrain feature. Therefore, this paper presents a modular approach to generate landscapes focused on the river deltas features. The modular proposal initially creates skeletons of deltas using a stochastic L-system grammar; we include the guidelines for the rules design. In the first module, we propose three L-systems that automatically create delta skeletons using these guidelines. The second module constructs the coastline and the sedimentary lands for the delta skeleton. Finally, the third module uses conditional generative adversarial networks (cGANs) to create the corresponding digital elevation models (DEMs) and land surface images. The evaluation of our proposal includes visual comparisons, and image quality metrics: the Frechet Inception Distance (FID), and the Naturalness Image Quality Evaluator (NIQE). The proposed modular integration generates realistic deltas with enough variability to outperform related work.

Highlights

Procedural Content Generation (PCG) is an area of study that aims to automate the creation of computing assets requiring limited human input
We propose an adaptation of Lindenmayer systems (L-systems) [10] to create skeletons of the river delta based on improvements of our previous work in [11]
The river delta structure is not enough to create a realistic landscape on its own; that is why we propose the integration of generative techniques such as conditional generative adversarial networks and L-systems

Summary

INTRODUCTION

Procedural Content Generation (PCG) is an area of study that aims to automate the creation of computing assets requiring limited human input. Our approach integrates an improved stochastic L-system to create delta structures with an image processing module to form their corresponding water map, activating generative models for its land surface image. It processes a considerable amount of data, such as water flow and soil features, and requires many iterations to simulate the deposition process Another method presented in [9], describes a fast and straightforward generation of rivers reaching the ocean. The improvement consists of a) using a new set of stochastic guidelines to create alternative succession rules, maximizing the probability of obtaining valid delta skeletons, and b) establishing ranges for the parameters variation, resulting in a more realistic emulation. Eliminate pairs of direction change symbols, as long as enough of these symbols remain within the system to maintain the sinuosity Using these guidelines to create rule variations, we have constructed three stochastic systems as valid examples. By following the guidelines it is possible to create L-systems with more rules and achieve more variability

THE WATER MAP MODULE

RESULTS AND DISCUSSION

CONCLUSION