There are several challenges for locust recognition, i.e., sample collection, small sample targets and multi-scale transformation in grassland locust images. A multi-scale grasshopper target detection and recognition model was proposed under complex background based on YOLO v5 network, which was used to recognize common grasshoppers in Ningxia grassland. To address the difficulty in sample collection, CycleGAN was used to expand the locust data set. Then, ConvNeXt was adopted to preserve the characteristics of small target locusts. Finally, Bi-FPN was utilized for neck feature fusion to enhance the capability of extracting locust features, which effectively solved the problem of large-scale transformation of locust photos. The experimental results showed that the best accuracy of the proposed model YOLO v5-CB was 98.6%, the mean average accuracy of the proposed scheme was 96.8%, and the F1 was 98%, which performed better than the Faster R-CNN, YOLO v3, YOLO v4 and YOLO v5. Using the improved model YOLO v5-CB, combined with the ecological environment collection equipment installed in Yanchi and Dashuikeng in Ningxia, a Web-based locust identification and detection platform was established, which had already been applied to grassland ecological environment data collection in Ningxia Yanchi Dashuikeng, Huangji Farm and Mahuang Mountain. This platform performed real-time tracking of locust in desert steppe of Ningxia, which can be further used for locust control in Ningxia. 