Procedural Terrain-Based Cloud Generation : Dataset Creation and Computational Analysis

Sammanfattning

Background: High-quality satellite image datasets are really important for training AI models used in remote sensing, land cover classification, environmental monitoring , and scene analysis. However , collecting balanced datasets that include different types of terrain under various weather and atmospheric conditions is challenging, costly and takes a lot of time. Real satellite images are often affected by issues like uneven cloud coverage, seasonal changes, and an imbalance of data across different regions. Objectives: The main objective of this thesis is to build an automated framework that can extract terrain-specific satellite images from NASA Landsat data covering all continents and improve these images by adding synthetic clouds and their corresponding shadows with adjustable settings. The goal is to create a scalable and diverse dataset that includes forests, grasslands,wetlands, mountains, deserts, urban regions and frozen areas for training and evaluating future AI models. Methods: The proposed system applies satellite image retrieval along with terrainbased filtering methods to gather images representing seven different terrain classes from various regions around the world. After that, a procedural generation pipeline is used to add clouds with varying densities, shapes, sizes and positions onto the images. Corresponding shadow masks are also created based on the placement of the clouds and certain lighting assumptions to enhance visual realism. The entire process is designed to be fully automated, allowing it to support large scale dataset generation. Results: The developed framework successfully generated a well structured dataset that includes multiple terrain categories combined with a wide range of cloud and shadow variations. The produced outputs helped increase environmental diversity while still maintaining the original visual features of each terrain type. The system demonstrated varying computational performance across terrain types, with execution time depending on terrain complexity. Conclusions: This thesis demonstrates that combining real satellite imagery with procedurally generated clouds and shadows is an effective way to create rich and diverse synthetic datasets. The proposed framework helps reduce the reliance on manually collected images with varying weather conditions and can support future research in remote sensing , computer vision and machine learning applications.