Mapping Conservation Practices via Deep Learning: Improving Performance via Hillshade Imagery, Sampling Design, and Centerline Dice Loss

The Iowa Best Management Practice (BMP) Mapping Project is a GIS database of conservation practices installed in Iowa’s fields as seen in aerial imagery from 2007 to 2010. In this study, we explore the feasibility of using convolutional neural networks (CNNs) to automate the process of image segmentation for several conservation practices in the database: grassed waterways, pond dams, terraces, and water and sediment control basins (WASCOBs). We experiment with imagery sources, sampling methods, transfer learning, neural network architectures, and loss functions to optimize segmentation performance. Our results demonstrate that lidar-derived hillshade imagery is important for identifying structural BMPs such as pond dams, terraces, and WASCOBs. Additionally, we show that a probability-proportional-to-size random sampling method for selecting training imagery outperforms CNNs trained on imagery sampled by systematic random sampling. We also find evidence that the centerline dice loss function helps to preserve the connectedness of linear BMP features. The results of this study show it could be feasible to develop an automated method of identifying BMPs from remote sensing imagery to monitor the adoption of BMPs across the Midwestern United States.