LFFNet: Layered feature fusion network enhanced passive infrared structural analysis in large-scale geomembrane covers

Abstract. Anaerobic lagoons at sewage treatment plants are covered with multiple sheets of high-density polyethylene (HDPE) geomembranes to prevent the emission of odorous gases and to harness biogas as a renewable energy source. Over time, raw sewage can accumulate and solidify into a mass, which can deform the covers and potentially affect their structural integrity. Currently, traditional passive thermography is limited by insufficient thermal excitations and has difficulty identifying features and substances beneath the covers. This study proposes a novel segmentation model, Layered Feature Fusion Network (LFFNet), to identify structural anomalies on large-scale geomembranes in sewage treatment plants by integrating infrared imaging and deep learning techniques. Notably, the model performs exceptionally well in addressing class imbalance issues, reaching a maximum Mean Intersection over Union (mIoU) of 94.68% and a maximum Mean Pixel Accuracy (mPA) of 97.49%.