This paper presents a real-time nonlinear laser fluorescence recognition method. First, the feature vectors consisting of transform coefficients were obtained by utilizing the three layers curvelet transform to decompose the pre-processing fluorescence spectrum of the heavy oil, diesel, crude oil and other types of common oils in various angles and different scales. Then the feature vectors were regarded as the parameters and sent into the support vector Machines (SVM) for training. Finally, the trained SVM was used for spectral classification of the oil slicks. Results from the trial suggest that it didn’t rely on a large number of samples, so that the number of support vectors was significantly reduced and the operation time was shortened for real-time running. Compared with traditional methods, the proposed method proves to be more efficient, faster and more reliable and has real-time capabilities.