<p>Accurate landslide detection is critical for geological hazard early warning, yet existing deep learning methods lack systematic architectural evaluation and inadequately account for morphological variations. This study proposed a multi-module synergistic architecture integrating Atrous Spatial Pyramid Pooling (ASPP), Squeeze-and-Excitation (SE) attention, and Path Aggregation Network (PANet) with an EfficientNet-B4 encoder for automated landslide segmentation from high-resolution satellite imagery. Comprehensive ablation experiments on the Bijie landslide dataset quantified individual module contributions: SE attention improved Intersection over Union (IoU) by 2.95%, ASPP contributed 2.75%, with complete integration achieving 90.06% IoU, 94.26% Precision, 95.18% Recall, and 94.70% Dice coefficient—surpassing state-of-the-art methods by 1.29–5.20%. We introduced continuous error metrics (MAE = 0.0116, MSE = 0.0085, R²=90.16%) computed directly on probabilistic predictions, enabling finer-grained discrimination when traditional binary metrics saturate.Our pioneering morphology-stratified analysis revealed that detection of circular landslides was significantly more accurate than that of long-strip landslides—the first systematic quantification of morphology-dependent performance in landslide detection. Mechanistic explanations—encompassing boundary complexity, resolution loss, and connectivity maintenance—provide theoretical foundations for targeted optimization. Attention heatmap analysis demonstrated more focused activation patterns with stronger background suppression for integrated architectures. The exceptional 95.18% Recall is critical for operational warning systems where missing landslides could have catastrophic consequences. This research delivers a deployment-ready model and transferable frameworks for diverse remote sensing segmentation tasks.</p>