Superpixel-refined deep learning framework with texture enhancement and multi-layer attention fusion for automatic crop detection in VGI cropland imagery

As a critical source of agricultural big data, images derived from volunteered geographic information (VGI) offer key advantages such as convenient acquisition, large data volume, precise geolocation, and high resolution. While serving dual roles as in-situ ground-truth validation for remote sensing-based crop mapping and supplementary datasets for satellite-derived monitoring, its processing challenges include spatial heterogeneity, enormous data scale, variable image quality, and inconsistent metadata. Existing methods struggle to simultaneously address texture ambiguity, contextual dependencies, and edge fidelity in diverse crop types. To tackle these issues, this study proposes a superpixel-refined deep learning framework integrating a texture-enhanced DeepLabv3+ architecture with multi-layer attention fusion. Specifically, we introduce a Gabor filter-based texture enhancement module to capture multi-scale textural details, addressing color homogeneity among similar crops. A hierarchical attention mechanism fuses positional and channel-wise dependencies across encoder layers, enhancing global context awareness for scattered crop distribution. Separable convolutions reduce computational overhead while maintaining accuracy. Superpixel-aided refinement via the simple linear iterative clustering (SLIC) and a voting-based ensemble further optimizes edge delineation, particularly for complex boundaries. Experimental results on the iCrop dataset demonstrate superior performance over classic networks, with a mean IoU of 87.76%, precision of 94.79%, and Kappa coefficient of 0.9133. The framework’s robustness stems from synergistic integration of texture priors, attention-driven context modeling, and edge-preserving optimization, establishing a foundation for real-world precision agriculture applications. Future work will focus on expanding dataset diversity to encompass various crop types and environmental conditions, while incorporating advanced computational techniques to enhance model robustness and generalization capabilities for real-world agricultural monitoring.