Development of a Machine Learning-Based Land Use Classification System Using Satellite Imagery

Thesis Abstract

Rapid urbanization and environmental change necessitate accurate and timely land use information to inform sustainable planning and resource management. Traditional land surveying methods are labor-intensive, costly, and often limited in spatial coverage and temporal resolution. Climate variability, urban sprawl, and land degradation further complicate land use monitoring, highlighting the urgent need for innovative, efficient, and scalable classification solutions leveraging advanced Information and Communication Technology (ICT). This study aims to develop a robust machine learning-based land use classification system utilizing high-resolution satellite imagery, with a specific focus on improving classification accuracy and operational efficiency. The primary objectives are to evaluate the effectiveness of different machine learning algorithms—namely Random Forest, Support Vector Machine, and Convolutional Neural Networks—in classifying land use categories, and to determine the optimal combination of spectral and spatial features for improved classification performance. Secondary objectives include assessing the impact of image pre-processing techniques such as normalization and segmentation and developing an integrated system that can be utilized by land use planners for routine monitoring. A quantitative research design is adopted, grounded in a positivist philosophy, emphasizing objective measurement and analysis. The population of the study comprises satellite images obtained from the Landsat 8 Thematic Mapper (TM) and Sentinel-2 MSI sensors covering a metropolitan region with diverse land use types, including residential, commercial, industrial, agricultural, and green spaces. A stratified random sampling technique was used to select 1,000 ground control points, ensuring representative distribution across categories. A total of 150 satellite image tiles, covering 500 square kilometers, will be collected, pre-processed, and annotated using existing land use maps and field validation data acquired through GPS-enabled ground surveys. Data collection instruments include high-resolution satellite imagery, GPS receivers for ground truthing, and geographic information system (GIS) tools for data annotation. The imagery will undergo radiometric and atmospheric correction, segmentation, and feature extraction, including spectral indices such as NDVI, SAVI, and built-up indices. Machine learning classifiers will be trained and validated using 70% of the annotated dataset, with the remaining 30% reserved for testing. Model performance will be evaluated through metrics such as overall accuracy, kappa coefficient, precision, recall, and F1-score, employing cross-validation techniques to ensure robustness. Additionally, statistical analyses such as repeated measures ANOVA will be used to compare the performance of different classifiers across land use categories. Expected findings include identification of the most effective machine learning algorithm for land use classification in the context of mixed land environments, with expected overall classification accuracies exceeding 85%. Incorporating spectral indices and texture features is anticipated to enhance classification precision, particularly in distinguishing complex categories such as mixed residential-commercial zones. The study also expects to demonstrate that convolutional neural networks outperform traditional algorithms in capturing spatial patterns due to their deep feature learning capabilities. The research contributes to the body of knowledge by providing an empirically validated framework for satellite image-based land use classification utilizing machine learning, facilitating more accurate, scalable, and cost-effective land monitoring systems. It bridges the gap between theoretical advancements in machine learning and practical applications in geo-informatics, offering insights into the integration of multi-source remote sensing data and automated classification techniques. In conclusion, the study affirms the potential of machine learning algorithms, particularly deep learning models, in enhancing land use classification accuracy and operational efficiency. It recommends further exploration of integrating multi-temporal data for dynamic land monitoring, and suggests the development of user-friendly decision support tools for land management agencies. Future research should focus on real-time classification systems and exploring the applicability across different ecological zones and urban contexts.