Large-scale　road　network　data　plays　an　important　role　in　fields　of　traffic　management,　urban　planning,　automatic　vehicle　navigation　and　emergency　management.　In　recent　years,　many　deep　learning　methods　have　been　applied　on　road　extraction　and　improved　the　accuracy　of　road　network　data.　However,　most　of　the　previous　studies　are　on　local　or　regional　scale,　and　it　is　still　a　challenging　task　to　accurately　extract　a　large-scale　road　dataset.　The　main　reason　for　it　is　the　influence　of　land　cover　type　diversity　on　the　accuracy　of　road　extraction.　In　this　paper,　Sentinel-2　data　will　be　used　for　large-scale　road　extraction　in　One　Belt　and　One　Road　area,　and　the　influence　of　different　land　cover　type　(city,　vegetation,　bare　soil)　on　road　extraction　is　comprehensively　analysed.　©　2021　ACM.