Forest encroachment prediction using multi-temporal satellite data and machine learning: a case study of Bandipur National Park, India

Abstract Forest encroachment poses a significant threat to protected forest ecosystems due to increasing human activities, including infrastructure development, agricultural expansion, and settlement growth. Continuous monitoring is therefore essential for effective conservation planning and sustainable forest management. This study presents an artificial intelligence-based framework for forest encroachment prediction in Bandipur National Park, India, using multi-temporal Sentinel- 2 satellite imagery integrated with topographic, land-cover, and anthropogenic variables. Multi temporal Normalized Difference Vegetation Index (NDVI) and Normalized Burn Ratio (NBR) were combined with elevation, slope, aspect, forest and water masks, and distance-to-road and distance-to-settlement variables. Three predictive models, namely Random Forest (RF), MultiLayer Perceptron (MLP), and a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) architecture, were comparatively evaluated. RF achieved the highest classification accuracy of 94.17%, followed by MLP at 91.50% and CNN–LSTM at 78.69%. Future projections for 2026 and 2030 indicated relatively limited encroachment under RF and MLP, with maximum projected areas of 1.27 km² (0.15%) and 0.64 km² (0.07%), respectively. In contrast, CNN–LSTM projected substantially larger encroached areas of 115.76 km² (13.24%) in 2026 and 71.94 km² (8.23%) in 2030. Overall, RF demonstrated the strongest and most consistent classification performance under the adopted experimental framework, indicating its potential for supporting forest monitoring, encroachment hotspot identification, and evidencebased conservation planning.