Abstract: [Purposes] Monitoring illegal rural construction on cultivated land remains a critical challenge due to land-cover heterogeneity and building scale variability.This study proposes an enhanced DeepLabv3+ framework with multi-scale feature fusion for accurate building extraction and suspected illegal patch identification.[Methods] Efficient channel attention (ECA)modules are embedded into the ResNet-50 encoder to recalibrate channel responses with negligible overhead.A feature pyramid network (FPN)performs top-down semantic propagation and lateral aggregation,enabling effective fusion of high-level semantics with spatial details.The atrous rates in ASPP are reparameterized to align the receptive field with small rural buildings.[Findings] On a subset of the WHU-Mix dataset,the method achieves 80.13% IoU and 88.97% F1-score,outperforming state-of-the-art baselines.Cross-scene validation using Henan-1 satellite imagery confirms the framework's robustness in cultivated land non-agriculturalization monitoring.[Conclusions] The proposed approach provides a reliable solution for building extraction,supporting natural resource supervision and enforcement at the grassroots level．

Key words: illegal rural construction on cultivated land, DeepLabv3+, ECA, FPN, cultivated land non-agriculturalization monitoring