The　critical　challenges　of　climate　change　have　made　carbon　emission　reduction　an　urgent　global　priority,　with　urban　transportation　systems　(UTS)　being　significant　contributors.　Rapid　urbanization　and　increased　traffic　demand　have　intensified　congestion　and　carbon　emissions.　However,　existing　traffic　signal　control　methods　often　rely　on　large　historical　data　and　pre-set　signal　timings,　making　it　struggle　to　adapt　dynamically　to　reduce　emissions　effectively.　This　study　proposes　a　novel　traffic　signal　control　method　based　on　Deep　Reinforcement　Learning　(DRL),　integrated　with　Cooperative　Vehicle-Infrastructure　Systems　(CVIS)　and　a　doubly　Day-to-Day　(DTD)　Dynamic　Traffic　Assignment　model,　aimed　at　improving　traffic　efficiency　and　reducing　carbon　emissions.　The　DTD　model　simulates　the　urban　road　network　as　the　training　environment,　while　CVIS　provides　real-time　traffic　flow　data　that　informs　the　DRL　model＇s　optimization　of　signal　timings.　Furthermore,　the　DTD　model　adjusts　traveler　departure　times　and　route　choices　for　significant　emission　reductions.　The　proposed　DRL-based　method　significantly　improves　signal　control　efficiency　and　carbon　emission　reduction　under　complex　traffic　conditions.　A　case　study　on　the　Sioux　Falls　network　indicates　that　the　DRL-based　traffic　signal　strategy　outperforms　traditional　fixed-time　control　strategies,　achieving　CO2　emission　reductions　of　21　%　to　27　%　in　various　scenarios,　particularly　excelling　in　the　S[sbnd]V　scenario.　Notably,　emissions　on　high-traffic　Link　28　(S[sbnd]V)　were　reduced　by　up　to　54.9　%.　This　study　underscores　the　potential　of　DRL　in　low-carbon　traffic　management　and　provides　practical　insights　for　the　sustainable　development　of　future　traffic　systems,　offering　robust　solutions　for　emission　reduction　and　efficient　traffic　management　in　UTS.　©　2025　Elsevier　Ltd