Evaluation of Advanced LULC Classification of Satellite Imagery with Convolutional Neural Networks

Abstract

Accurate Land Use/Land Cover (LULC) classification plays a crucial role in sustainable environmental management, urban planning, and agricultural monitoring. Traditional classification techniques often struggle with the complexity of heterogeneous landscapes, particularly in rural areas like Ekiti West Local Government Area, Nigeria. This study employs a Convolutional Neural Network (CNN) framework for LULC classification using Sentinel-2 multispectral satellite imagery combined with ground-truth data. The methodology includes comprehensive data pre-processing, image augmentation, CNN model development, and hyper-parameter optimization. The CNN architecture was specifically designed to handle multispectral inputs and spatial variability. Results show that farmland dominates the landscape, followed by forest, bare land/rock, and built-up areas. The CNN model achieved high classification performance with an overall accuracy of 96.86%, demonstrating strong precision, recall, and F1-scores across all classes. This study confirms the potential of CNN-based approaches for reliable and scalable LULC mapping in data-scarce rural environments and contributes to the growing body of geospatial Artificial Intelligence applications in Africa.