Application of spectral indices and deep learning (convolutional neural network model) on land cover change analysis

Background: Understanding land cover change is crucial for sustainable urban development, particularly in rapidly growing coastal cities such as Semarang City, Central Java, Indonesia. Methods: This study investigates spatial and temporal patterns of land cover change from 2000 to 2025 by integrating multi-temporal Landsat satellite imagery, key spectral indices—namely the normalized difference vegetation index, normalized difference water index, and normalized difference built-up index—and a deep learning approach based on convolutional neural networks. Annual Landsat images were preprocessed for atmospheric correction, cloud masking, and spatial subsetting using Google Earth Engine. Adaptive thresholding was then applied to each spectral index to delineate vegetation, water bodies, and built-up areas. Findings: Quantitative analysis revealed a significant decline in vegetation cover, with the normalized difference vegetation index dropping from 53.66% (397.59 km²) in 2000 to 46.83% (346.98 km²) in 2025, driven by urban expansion and landscape conversion, especially in coastal and lowland areas. Normalized difference water index analysis indicated a reduction and fragmentation of water bodies after 2015, linked to reclamation, sedimentation, and urban encroachment. Conversely, built-up areas expanded steadily, confirming accelerated urbanization. Scatter plot and regression analyses showed strong inverse relationships among vegetation, water, and built-up land, emphasizing ecological trade-offs and the loss of green-blue infrastructure. Conclusion: To enhance classification accuracy, a convolutional neural network was trained and validated on image patches, achieving a validation accuracy of 60%—outperforming conventional threshold-based methods by better capturing complex spatial patterns. The integrated remote sensing and deep learning framework offers robust potential for long-term, large-area land cover monitoring. Novelty/Originality of this article: The novelty of this research lies in its combined use of spectral indices and deep learning for multi-decadal land cover change analysis, providing a transferable methodology for other rapidly urbanizing coastal cities.