Detecting subsurface drainage pipes using a fully convolutional network with optical images

More than half of croplands in the Midwestern United States are equipped with subsurface drainage pipes to reduce excess water in productive but wet areas. The use of drainage systems not only reduces subsurface water table to prevent waterlogging and flooding but also increases nutrient losses by developing artificial preferential flow paths. The exact locations of subsurface drainage pipes are thus imperative to manage and monitor water quality and nonpoint source pollution. However, such data are not widely available due to private ownership. Previous studies used conventional image filtering methods, thermal images, and ground penetration radar to detect subsurface drainage pipes. Due to surface features, such as furrow and depressions, and their limited data availability, these experiments did not provide a robust approach to identify subsurface drainage pipes over a large area. To overcome these limitations, in this study, we propose a subsurface drainage pipe detection approach based on deep learning with optical images. Our deep learning approach uses a fully convolution network (FCN) architecture that takes an optical image patch as an input and gives an output of pixel-wise drainage pipe detection map. The FCN was trained and validated using optical image datasets obtained from a freeware Google Earth that provides temporally and spatially abundant data. The trained FCN was then applied to large-scale drainage pipe detection tasks to evaluate its performance. The performance comparison between the proposed deep learning approach and conventional image processing techniques (Sobel and Canny edge detection methods) was also carried out. The results demonstrate that the proposed deep learning approach shows accurate and robust drain line detection performance with an average Dice coefficient of 0.58 for validation sets, providing superior performance over the conventional image processing techniques.