Abstract
Land Cover and Land Use (LCLU) segmentation plays a fundamental role in various remote sensing applications, including environmental monitoring, urban planning, and disaster management. Traditional models often face limitations in real-time processing and deployment on resource-constrained devices due to their high computational requirements. This paper presents a lightweight neural network designed to address these challenges by integrating dense dilated convolutions with pyramid depthwise convolutions for multiscale feature extraction. The proposed encoder-decoder architecture utilizes dense connections to aggregate spatial and contextual information across different resolutions, enhancing segmentation accuracy while minimizing computational overhead. The model's performance was rigorously evaluated using the NITRDrone and UDD6 datasets, demonstrating a segmentation accuracy of 94.8%, with a significantly reduced parameter count compared to state-of-the-art methods. The compact design of the network facilitates its implementation on low-power devices, enabling real-time LCLU analysis across diverse environmental conditions. This work underscores the potential of lightweight neural networks to advance remote sensing image processing, offering scalable and efficient solutions for practical applications in geospatial analysis.