Evaluation of Deep Learning Benchmarks in Retrieving Outfalls Into Rivers With UAS Images

Outfall surveys are critical to basin environmental and ecosystem protection. High-spatial-resolution remote sensing has been a key technology in outfall surveys. However, manual visual interpretation of outfalls from unmanned aircraft system (UAS) images is inefficient. Recently, the convolutional neural network (CNN) deep learning method has been demonstrated to be an efficient and accurate method for retrieving specific objects from remote sensing imagery. However, CNN’s architecture and features are often proposed to detect common objects from universal images that are less complex than the specific remote sensing objects. Thus, an optimal CNN benchmark is significant to the performance of CNN-based approaches in the field of remote sensing. In this study, we evaluate the performance of several state-of-the-art CNN architectures and features for improving the geo-deep learning CNN for outfalls (GDCNN-outfalls) we proposed. We found that Faster R-CNN combined with feature pyramid network (FPN) and generalized intersection over union (GIoU) or complete intersection over union (CIoU) are the most suitable benchmarks, with ${AP}_{{50}}$ values of 68.7 and 67.0, respectively, which are better than those of cascade R-CNN ( ${AP}_{{50}}$ of 64.6) and Libra R-CNN ( ${AP}_{{50}}$ of 62.7). You only look once (YOLO)-v4 can be an alternative with an ${AP}_{{50}}$ of 66.8 and comprehensively considers retrieval accuracy, the current GPU performance, and the application needs. This result indicates that the latest CNN models in the field of computer vision are not always the most suitable benchmarks for specific object retrievals from remote sensing imagery. Our study also provides a typical workflow for the relevant CNN-based model designs and applications.