CNN-driven recognition of fluvial terrace sequences along the middle Yangtze River and its neotectonic implications

Fluvial terraces preserve critical records of Quaternary tectonic uplift and river incision, yet their regional-scale identification remains hampered by the subjectivity and spatial discontinuity of conventional methods. This study presents a convolutional neural network (CNN) framework, built on a modified U-Net encoder-decoder architecture, for automated, multi-level terrace sequence recognition along the approximately 600-km Yichang-to-Wuhan corridor of the middle Yangtze River. The model ingests six-channel DEM-derived morphometric inputs-elevation, slope, profile curvature, planform curvature, topographic wetness index, and topographic position index-computed from the ALOS AW3D30 DEM, with training labels derived from 46 RTK-GNSS-surveyed terrace cross-sections. Spatial block cross-validation was employed to partition the dataset, ensuring geographic isolation between training and test tiles. Evaluated against independent test data, the CNN achieves an overall accuracy of 92.3%, a Kappa coefficient of 0.896, and a macro-averaged F1 score of 89.0%, outperforming slope-threshold, object-based image analysis, and Random Forest benchmarks-all of which underwent equivalent hyperparameter optimization-particularly for higher, more fragmented terrace levels. Four terrace levels (T1-T4) were mapped across the study corridor, revealing a systematic west-to-east decline in terrace completeness that mirrors the regional tectonic gradient from the Huangling anticline to the subsiding Jianghan Basin. Quantitative analysis of terrace deformation yields tectonic uplift rates ranging from 0.28 mm/year near Yichang to 0.06 mm/year in the basin interior, with localized elevation offsets of 8-15 m across major fault zones. A tectonic activity intensity index (TAI), supported by spatial buffer statistics, highlights the Yichang and Wuhan sub-reaches as the most actively deforming segments. We acknowledge that climatic oscillations modulate terrace formation and that the current framework cannot fully decouple tectonic from climatic signals; nonetheless, the spatially coherent deformation patterns localized along fault traces favor a predominantly tectonic interpretation. These findings demonstrate that CNN-assisted terrace mapping can provide spatially continuous neotectonic insights unattainable by field methods alone.