Satellite imagery and deep learning framework for highway construction with minimal vegetation impact

Abstract

Constructing roads through vegetated terrain forces a trade-off between minimizing ecological harm and preserving constructional ease. This work presents a unified, end-to-end framework that reconciles both objectives, with three major contributions. First, propose a hybrid perception model that fuses a graph convolutional network (GCN)—capturing scene-level spatial context—with a capsule network (CapsNet) that encodes part-whole relationships for reliable detection of small, clustered vegetation; this configuration attains an average precision (AP) of 0.925, average recall (AR) of 0.948 (intersection over union (IoU) = 0.5), and 98% land-cover classification accuracy on unseen data. Second, we introduce an ecological-sensitivity cost map fusing vegetation normalized difference vegetation index (NDVI) with terrain slope, over which we evaluate Dijkstra, A*, a novel Eco-Aware A* embedding a vegetation-avoidance penalty, and an non-dominated sorting genetic algorithm II (NSGA-II) multi-objective search that jointly minimizes path length and vegetation damage. Third, on a benchmark modeled on the Bengaluru–Kanakapura road (NDVI = 0.7, slope = 0.3), the eco-sensitive route lowers segment NDVI by roughly 69% and eliminates all high-canopy exposure while adding only about 8% to path length versus an optimized non-vegetated route. Together, these components enable automated, ecologically aware road planning directly from imagery.