An Improved Deep Learning Approach for Retrieving Outfalls Into Rivers From UAS Imagery

Abstract

Outfalls into rivers are the final gate of anthropogenic pollution flowing to receiving waters, which means that outfall surveys are significant to basin environmental protection and ecosystem health management. Unmanned aircraft systems (UAS) with high spatial resolution imagery have become important data for ongoing surveys of outfalls. However, outfalls retrieval from UAS imagery is inefficient to visual interpretation and a challenging task for traditional spectral-based and object-oriented classification methods given the problems of salt-and-pepper noise and scale selection. In this study, an improved geo-deep learning approach based on the faster region convolutional neural network (R-CNN) architecture (GDCNN-outfalls) is proposed for retrieving outfalls into rivers with UAS imagery. In the proposed method, three tactics—anchor size, region of interest (RoI), and hard negative mining—were adopted to optimize the benchmark Faster R-CNN application in outfalls retrieval. Meanwhile, a geo-classifier module with digital surface model (DSM) enhancement and a spatial activation function was integrated with the Faster R-CNN architecture to generate GDCNN-outfalls. The validation experiments indicated that GDCNN-outfalls improved the performance of Faster R-CNN in outfall retrieval by suppressing false positive (FPs) from 33.52% to 26.14% and increasing the F1 score from 0.72 to 0.75. The test results confirm the performance of GDCNN-outfalls with a recall of 79.3% and higher precision (48.4%) than that of Faster R-CNN (2.1%), also show the GDCNN-outfalls is ten times faster than visual interpretation. This study demonstrates that the combination of deep learning and UAS techniques can be a feasible solution to detect outfalls in outfall surveys.