An adaptive recognition method for crop row orientation in dry land by combining morphological and texture features

Accurate identification of crop row orientation (CRO) is crucial for agricultural management . Most current CRO identification methods rely on image texture features from very-high-resolution (VHR) images, but their recognition accuracy still remains challenging, especially for large-scale mapping. To achieve rapid, cost-effective, and accurate large-scale CRO identification, an adaptive method was proposed. Vector cropland parcels generated on a cloud platform were combined with VHR imagery to adaptively identify CRO based on morphological and texture features. The effectiveness of the adaptive method was validated at Youyi Farm, Heilongjiang Province. The results are as follows: (1) A total of 4159 dry cropland parcels were extracted after removing paddy fields and a few non-cropland regions using the Normalized Difference Water Index (NDWI) and the Ratio Vegetation Index (RVI). The mean Intersection over Union (mIoU) was 70.5 %, and the EP (Extraction Precision) was 0.88, indicating that the overall parcel morphology generally aligns with the actual parcel shape. (2) By adjusting the parcel Length-to-Width ratio (L/W) to balance the CRO Recognition Rate (RR), Precision (Prec), and Root Mean Square Error (RMSE), an optimal L/W of 1.4 was determined, achieving the best overall balance. (3) Under the optimal L/W, the morphological feature method demonstrated a lower identification rate (RR: 67.3 %) but higher accuracy (Prec: 89 %) with a lower deviation (RMSE: 23.6°), while the texture feature method showed the opposite trend (RR: 89.4 %, Prec: 68 %, RMSE: 36.9°). Combining both features significantly improved the identification rate (RR: 94.7 %) while maintaining a low deviation (RMSE: 25.75°), indicating that the adaptive CRO identification method achieves optimal performance. The proposed method enables rapid and accurate CRO identification, supporting regional-scale CRO mapping. • Cloud platform and deep learning boosted land boundary extraction accuracy. • A threshold of parcel L/W was optimized for balancing the stability and accuracy. • New CRO method enhances robustness with texture and morphological features.