Green AI for Sustainable Transportation Infrastructure: A Sys-tematic Review of Energy-Efficient Deep Learning in Railway, Highway, and Smart Mobility Systems (2020-2026)

Source abstract

AI in transportation is built largely on deep learning models. The models run on edge devices with hard energy constraints. To be green, they must therefore be energy efficient. Much of the literature labels them "green". This review examines whether such claims rest on direct sustainability evidence; to our knowledge, none has been published. Following PRISMA 2020, we searched Scopus, IEEE Xplore, and Web of Science (2020-June 2026). We included 721 studies applying Green AI techniques: pruning, quantization, knowledge distillation, lightweight architecture, TinyML, and dedicated accelerators. They cover roads and ADAS, railway, connected and autonomous vehicles, and intelligent sensor networks. Each study was classified by its strongest evidence: a direct sustainability metric (energy, power, power efficiency, battery life, CO2) or computational proxies. In our results, 58 studies (8.0%) report a direct metric. One study reports a carbon figure. Most of the 58 address vehicle-centered applications. Railway contributes 3 studies. 48 of the 58 report power or energy values measured on the target hardware, not modelled. The largest measured saving: 1,961.8-times lower energy per inference than a CPU baseline. We therefore recommend to authors, editors, and reviewers in transportation AI that every efficiency claim include at least one direct metric on a named target platform.