Multi-Scale Feature Learning for Farmland Segmentation Under Complex Spatial Structures.

Area of Science:

• Computer Vision
• Remote Sensing
• Agricultural Science

Background:

• High-resolution remote sensing imagery of farmland presents spatial complexity due to fragmented parcels.
• Boundary ambiguity and spectral confusion hinder effective feature discrimination in semantic segmentation tasks.
• Existing methods struggle with the intricate details of agricultural landscapes.

Purpose of the Study:

• To develop a novel semantic segmentation model, CSMNet, for complex agricultural landscapes.
• To improve feature discrimination and boundary delineation in remote sensing imagery.
• To address challenges of scale heterogeneity and class imbalance in farmland parcel segmentation.

Main Methods:

• Utilized a ConvNeXt V2 encoder for hierarchical representation learning.
• Implemented a multi-scale fusion architecture with enhanced skip connections and lateral outputs.
• Incorporated an adaptive multi-head attention module for dynamic contextual cue integration.
• Employed a hybrid loss function (Binary Cross-Entropy and Dice loss) to manage class imbalance.

Main Results:

• CSMNet achieved high performance metrics: Precision (95.91%), Recall (93.95%), F1-score (94.92%), and IoU (90.85%).
• The model significantly outperformed state-of-the-art methods, including Unet++, PSPNet, SegNet, DeepLabv3+, TransUNet, SeaFormer, and SegMAN.
• Demonstrated superior IoU compared to Unet++ by 8.92% and other methods by over 2% to 15%.

Conclusions:

• CSMNet effectively improves information utilization and boundary delineation in complex agricultural remote sensing imagery.
• The proposed model shows significant advancements in semantic segmentation for fragmented and scale-heterogeneous farmland parcels.
• CSMNet offers a robust solution for precise agricultural landscape analysis.