With　the　increasing　demand　for　dynamic　control　of　sewage　quality,　developing　a　fast　and　accurate　detection　method　that　can　continuously　monitor　total　organic　carbon　(TOC)　in　water　has　become　a　crucial　issue.　However,　traditional　methods　for　TOC　detection　require　a　high　consumption　of　time　and　resources.　To　overcome　these　drawbacks,　in　this　paper,　a　rapid　and　effective　TOC　detection　method　is　developed　based　on　near-infrared　spectroscopy　(NIRS)　and　an　improved　convolutional　neural　network　(CNN).　Firstly,　the　spectrum　of　the　water　samples　is　collected　using　a　portable　near-infrared　spectrometer.　Then　a　Convolutional　Block　Attention　Module　(CBAM)　is　integrated　with　one-dimensional　CNN　to　effectively　extract　key　characteristics　of　pollutants　and　achieve　accurate　quantitative　analysis.　The　feature　visualization　results　demonstrate　the　effectiveness　of　the　feature　extraction.　And　comparison　experimental　results　shown　that,　the　proposed　method　model　has　the　greatest　coefficient　of　determination　and　the　smallest　root　mean　square　error　compared　with　other　models.　Thus,　the　proposed　method　can　accurately　detect　the　total　organic　carbon　in　water　and　is　suitable　for　practical　applications.　©　2023　IEEE.