AI-Driven Highway Route Planning for Minimal Vegetation Loss

Abstract

Satellite remote sensing has transformed the ability to characterise land cover, vegetation density, and terrain conditions over wide geographic extents at high spatial resolution. This paper presents a framework that integrates multi-band satellite imagery with machine learning to identify optimal highway routes with minimal vegetation loss. A synthetic 600×600-pixel multi-spectral scene (calibrated to Sentinel-2/Landsat-8 reflectance ranges), representing a 6×6 km study area, is processed to extract six spectral indices: NDVI, NDWI, NDBI, EVI, BSI, and Land Surface Temperature (LST). A Random Forest classifier trained on 12,000 pixel samples achieves 99.5% accuracy in mapping eight land cover classes. Satellite-derived features from 10,000 road segments are used to train a Gradient Boosting Regressor (GBR), which achieves R² = 0.9798 for vegetation loss score prediction. The GBR predictions serve as edge weights in an eco-weighted directed terrain graph (2,310 nodes; 17,658 edges); Dijkstra-based eco-routing identifies alignments that reduce average vegetation loss by 6.9% relative to a cost-only baseline, while the hybrid route simultaneously reduces construction cost by 0.8% and vegetation impact by 6.8%. The framework thus enables a proactive reduction of ecological impact in highway alignment design.