The level of informationization infrastructure of the power system is constantly improving, and it is of great practical significance to carry out real-time perception and early warning of environmental risks during the construction period of the project based on image processing algorithms. This paper proposes a multi-scale parallel real-time detection algorithm based on SSD, which optimizes the network structure of SSD algorithm, combines and splices different sizes of inverted residual blocks and different types of activation functions with each other, and designs a kind of lightweight feature extraction network EPNets. Then, it proposes a lightweight parallel fusion structure, which is applied to the multi-scale prediction of the lightweight feature extraction network, and optimizes the environmental risk real-time detection speed of the algorithm. The algorithm is optimized for realtime environmental risk detection speed. A Bayesian network-based environmental risk behavior warning model is constructed to provide real-time warning for the detected risk behaviors. By comparing with the original algorithm and existing target detection algorithms, the multi-scale parallel fusion detection algorithm based on SSD proposed in this paper can maintain good detection speed with low loss degree, and its environmental protection risk identification time is only 9ms.Meanwhile, the early warning algorithm in this paper realizes the accurate early warning of the soil erosion risk in the study area through the soil erosion environmental protection risk during the construction period of transmission and transformation projects detected. It provides an objective guideline for the control of environmental protection risks and work priorities.