The invention discloses a near-ultraviolet image processing based detection and early-warning method for hazardous chemical substances in an off-shore marine site. The method comprises a training preparation stage and a detection and early warning stage. The training preparation stage comprises obtaining and synthesizing training images, preprocessing and marking the images, training a model and preparing a panoramic map. The detection and early warning stage comprises obtaining a detection image, preprocessing the detection image, detecting and determining a target area, GPS matching, image fusion and graded early warning. A rapid deep neural network confidence detection model is established for the hazardous chemical substances in the near-ultraviolet image, the fact that the reflectivity of a colorless hazardous chemical substance is obviously different from that of water in the near ultraviolet band is utilized fully, and the difficulty in carrying out imaging detection on the floating hazardous chemical substance due to colorlessness is overcome; and GPS information matches the panoramic map for the image which is determined to include the floating hazardous chemical substance, the images, including the hazardous chemical substances, with adjacent sequential numbers are spliced and fused, much fewer images with the hazardous chemical substances need to be processed, and the method is characterized by being rapid, accurate and high-efficiency.