UniRepLKNet-SE With Cross-Scale Feature Fusion for Abandoned Cropland Mapping

Source abstract

Accurate mapping of abandoned cropland supports food-security assessment, land-restoration planning, and regional carbon accounting, yet early abandonment remains difficult to distinguish from temporary fallow, low-input cultivation, grass regrowth, and seasonally exposed soil. This study develops UniRepLKNet-SE, a dense remote-sensing interpretation model that combines large-kernel representation learning, squeeze-and-excitation channel recalibration, and an attentive cross-scale decoder. A controlled, internally consistent benchmark was assembled from multiseason Sentinel-2 indices, Sentinel-1 backscatter, terrain attributes, and parcel-context variables on a 10 m grid. Evaluation used 18,420 annotated patches from four agricultural landscape strata with spatially blocked training, validation, and test subsets. UniRepLKNet- SE achieved 93.8% overall accuracy, 89.6% mean F1, 78.4% abandoned-cropland intersection over union, and 76.5% boundary F1. Target-class intersection over union exceeded Swin-UNet by 4.8 percentage points and DeepLabV3+ by 7.3 points. Cross-scale fusion increased boundary F1 from 71.2% to 76.5%, while channel recalibration reduced false abandonment alarms in sparse winter-wheat fields by 13.6%. Ablation and missing-season tests indicated that wide spatial support contributed most strongly in fragmented terraces, whereas multiresolution fusion stabilized predictions under phenological mismatch. The results support the proposed architecture as a reproducible engineering candidate for parcel- oriented screening and verification prioritization; external geographic validation and multi-year management evidence remain necessary before policy-level deployment.