| Abstract [eng] |
Advances in optical satellite technology and reduced launch costs have fueled demand for geospatial intelligence. The dissertation explores the transformative impact of satellite imagery on global economic event prediction and addresses challenges in computer vision research, focusing on accuracy and prediction speed, critical for time-sensitive applications. It proposes enhancing object recognition model design “Sat-mod” with training, and complexity regulation. A framework optimizes a Fully Convolutional Neural Network (FCN) for fast object recognition in optical satellite imagery. Due its light computational architecture (6.9832 G-FLOPs), the FCN (UNET) excels in speed and performance, surpassing existing networks in light-vehicle object recognition. It offers a fivefold improvement in training time and a sevenfold faster network-specific prediction, essential for real-time applications like algorithmic trading. The dissertation also addresses the object prediction accuracy limitations of manually-designed neural networks in complex and dispersed multi-spectral satellite imagery datasets. It highlights unique dataset properties and proposes Automated Machine Learning (AutoML) techniques and the Neural Architecture Search framework NAS-MACU, which automatically designs neural network architecture adapted to solve dispersed problem ranges. NAS-MACU proved highly effective in low-information settings achieving the F1 score of 0.934 and surpassing manually-design networks. |
