Analyzing Multiclass Land Cover and Spatial Point Patterns on Sentinel-2 Imagery Using Machine Learning and Deep Learning

Land cover conversion around educational centers, such as universities, is an inevitable consequence of increasing urban activity. The development of boarding houses, commercial zones, and other infrastructure often follows the expansion of academic institutions. To support sustainable spatial planning, early identification of land cover and analysis of spatial distribution patterns are crucial for zoning regulation and infrastructure management. This study focuses on classifying land cover and analyzing spatial patterns around Universitas Riau (UNRI) using Sentinel-2 satellite imagery with a 10-meter spatial resolution. The research applied a supervised classification approach, utilizing spectral bands—specifically Near-Infrared (NIR) and Short-Wave Infrared (SWIR)—as explanatory variables. The response variable was land cover, categorized into vegetation, non-vegetation, and water. Three machine learning models—Support Vector Machine (SVM), Naïve Bayes (NB), and Backpropagation Neural Network (BNN)—were compared based on overall accuracy and the Kappa coefficient. The models were trained and tested using a stratified 80-20 data split to ensure a balanced evaluation. Among the models, SVM demonstrated the highest accuracy, achieving an average of 91.15% in 2022 and 83.90% in 2023 with minimal variance, confirming its reliability for land cover classification. Spatially, non-vegetation areas were concentrated near major access routes and facilities, highlighting the influence of infrastructure development on land conversion. The study also identified potential growth zones within a 3–5 km radius from UNRI, emphasizing the need for anticipatory and sustainable land use policies. These findings support the formulation of spatial strategies aligned with Law No. 26 of 2007 on Spatial Planning and offer valuable insights for guiding urban development around academic hubs.