Water　quality　detection　is　very　important　for　monitoring　water　sources　and　main　canal,　which　is　beneficial　to　offer　strategies　for　the　management　of　water　quality　and　environment.　According　to　the　practical　distribution　and　data　characteristic,　this　paper　proposes　a　semi-supervised　detection　method　of　water　quality　based　on　a　sparse　autoencoder　network.　In　the　proposed　approach,　an　IoT-based　distributed　structure　is　implemented　to　execute　data　interaction,　and　a　representation　model　is　firstly　learned　via　a　sparse　autoencoder　trained　by　unlabeled　water　monitoring　data　acquired　from　8　physical　reservoirs,　then　a　softmax　classifier　is　trained　using　a　small　set　of　labeled　classification　data　based　on　the　China　Surface　Water　Environmental　Quality　Standard　(GB3838-2002)　expressed　by　the　sparse　autoencoder.　The　combined　model　is　finally　used　to　evaluate　the　water　quality.　Compared　Experimental　results　with　the　traditional　methods　and　actual　measure　results　show　that　the　proposed　method　has　high　robustness　and　accuracy　for　water　quality　assessment,　and　has　a　good　prospect　of　practical　applications.　©　2016.