Cooling tower emissions pollute the atmosphere. Using high-resolution remote sensing images to detect cooling towers can provide decision-making data for the treatment of exhaust emissions. Aiming at the problems such as low detection accuracy and slow detection speed of traditional algorithms in high-resolution remote sensing image object detection, the authors improved the RetinaNet by adopting a sampling-free mechanism to detect the cooling towers. First, Images in dataset were labeled as working cooling towers and resting cooling towers. Then, based on the number of object categories in the dataset and the proportion of positive samples in training, the bias term of the last layer in the classification subnetwork and class-adaptive threshold were determined. In addition, the regression loss was used to determine the adjustment ratio of the classification loss to avoid loss functions to be dominated by numerous background examples. Finally, ResNet50 was used to extract image features, and the FPN module was used to generate a multi-scale convolution feature pyramid. Detection boxes regression and category confidence calculations were performed for each layer of features. The results show that, for cooling tower detection on high-resolution remote sensing images, the proposed algorithm can improve the detection accuracy while ensuring the detection speed compared with RetinaNet, which proves the effectiveness of the proposed algorithm.