What question did this study set out to answer?

To improve accuracy in mapping cultivated lands in Northeast China's Black Soil region using remote sensing imagery and advanced modeling techniques.

March 26, 2026Open Access

Fusion of residual networks and hybrid attention mechanism for high accuracy cultivated land mapping in northeast china’s black soil region using planet imagery

Key Points

To improve accuracy in mapping cultivated lands in Northeast China's Black Soil region using remote sensing imagery and advanced modeling techniques.
Used Planet remote sensing imagery for data collection.
Employed an enhanced RC-UNet model with ResNet backbone for semantic segmentation.
Integrated a hybrid channel-spatial attention mechanism to improve local detail refinement and accuracy.
Achieved 96.87% accuracy in farmland boundary delineation, a 3.09% increase over standard U-Net.
Outperformed various models including SwinUNet, TransUNet, and DeepLabV3+ by significant margins.
Achieved an Intersection over Union (IoU) of > 90% and < 3% area estimation error.

Abstract

Cultivated lands in Northeast China exhibit extensive spatial distribution with pronounced regional heterogeneity in both soil types and cropping patterns. Although exhibiting gentle slopes overall, certain regions are characterized by prolonged slope lengths coupled with frequent freeze-thaw cycles and pronounced wind/water erosion, collectively enhancing spatiotemporal heterogeneity in surface vegetation cover and soil properties. These natural processes significantly increase spectral confusion between cultivated and noncultivated areas in remote sensing imagery, challenging conventional pixelbased or shallowfeature classification approaches in both accuracy and computational efficiency. Such limitations fundamentally constrain their applicability for highaccuracy cultivated land mapping. This study focuses on Youyi County, Shuangyashan City, using Planet remote sensing imagery and an enhanced RC-UNet semantic segmentation model for farmland parcel extraction. During model training, the U-Net backbone was replaced with ResNetto alleviate gradient vanishing and improve information transfer. Additionally, a hybrid channelspatial attention mechanism was incorporated to refine local details and boost segmentation accuracy. Results demonstrate that the improved RC-UNet model significantly enhances farmland boundary delineation, achieving 96.87% accuracy—a 3.09% increase over the standard UNet. Compared to SwinUNet, TransUNet, Unet++, DeepLabV3+, PSPNet, SVM, and RF models, it outperforms them by 1.65%, 2.62%, 2.28%, 3.13%, 5.44%, 6.55%, and 6.2%, respectively. The enhanced RC-UNet model demonstrates superior feature extraction capabilities, achieving an Intersection over Union (IoU) of > 90% in field boundary delineation. This advancement enables precise agricultural monitoring with < 3% area estimation error, which is critical for ensuring regional food security and promoting sustainable intensification practices.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper