Abstract: Artificial intelligence has rapidly advanced in smart oilfield applications. However, efficient and accurate methods for inspecting gas pipeline trench backfill quality remain lacking. This results in low inspection efficiency and accuracy, causing issues such as backfill soil settlement and uneven pipeline stress. To address these challenges, an intelligent evaluation method was proposed for trench backfill quality based on semantic segmentation. Specifically, a Residual Feature Fusion Attention Segmentation Network (RFFA-SegNet) was developed to extract deep features of debris such as crushed stones and bricks using a residual structure. The integration of a fusion attention mechanism further enhanced the model's ability to delineate debris edges. Additionally, a transfer learning strategy was employed to reduce training time and improve the model's generalization. Experiments on a field-collected dataset of 1 060 images demonstrated that the proposed algorithm achieved an accuracy of 99.65% and a processing speed of 23.08 frames per second (FPS). Compared to the baseline, accuracy improved by 3.02% and FPS by 11.71, outperforming most mainstream segmentation models. The proposed method significantly addresses traditional detection issues such as blurred edges, target adhesion, false positives, and false negatives. It enables accurate and rapid segmentation of backfill debris, providing reliable technical support for particle size measurement and quantitative analysis of debris uniformity. This approach offers a novel solution for intelligent evaluation of gas pipeline trench backfill quality and serves as an effective reference for feedback control in excavator robot construction.