Abstract

The extraction of meaningful patterns from topographical imagery has immense applications in geospatial analysis, environmental monitoring, and urban planning. However, existing methods often struggle with scalability and real-time adaptability. Traditional approaches rely heavily on handcrafted features, limiting their ability to generalize across diverse terrains. These methods are computationally intensive and fail to leverage modern deep learning capabilities for robust pattern recognition. This study proposes MobileVNet, a lightweight deep learning model designed for efficient geospatial data mining. MobileVNet employs a hybrid encoder-decoder architecture, integrating convolutional blocks optimized for edge devices. Using a dataset of 10,000 topographical images, MobileVNet was trained to classify and segment patterns like ridges, valleys, and water bodies. MobileVNet achieved an accuracy of 94.6%, surpassing state-of-the-art models like U-Net (92.1%) and SegNet (90.5%). It reduced inference time by 35%, making it suitable for real-time applications.