Abstract:

Objective: To address the shortcomings of existing remote-sensing extraction methods of farmland shelterbelts in linear structure expression, spatial continuity maintenance, and regional generalization ability, this study proposes an automatic extraction method (ENVINet5-SME) integrating ENVINet5 and Structural Morphological Enhancement (SME) to improve the recognition accuracy, spatial-structure expression ability, and model transferability. Method: With the 3 m resolution multispectral images of Jilin-1 (JL1GF02A) as the data source, a sample dataset was constructed for three typical shelterbelt regions in Changchun City with significantly different site conditions. First, NDVI and LBP texture features were calculated and combined with NIR and RGB bands to form four feature combinations: NIR+RGB, NIR+RGB+NDVI, NIR+RGB+LBP, and NIR+RGB+NDVI+LBP. Training and classification were performed based on ENVINet5. To address the problems of noise patches and fragmented boundaries in the classification results, a morphological enhancement process of “median filtering-erosion and dilation-thinning” was introduced to optimize spatial connectivity and geometric structure. To evaluate the classification performance of the proposed method, deep learning models such as DeepLabV3 and FPN, as well as the machine learning model Random Forest, were adopted as comparative approaches. F1-score, Producer’s Accuracy (PA), User’s Accuracy (UA), and mean Intersection over Union (mIoU) were used for accuracy assessment. Furthermore, transfer learning experiments were conducted with Study Area Ⅱ as the source domain to validate the model’s transferability and training efficiency. Result: Morphological enhancement effectively reduced noise patches and restored fragmented shelterbelt structures, thereby improving the connectivity and integrity of shelterbelt spatial patterns. Comparative experiments demonstrated that ENVINet5 achieved satisfactory classification performance. After incorporating Structural Morphological Enhancement (SME), the F1 score and mean Intersection over Union (mIoU) of ENVINet5-SME further increased to 0.7439 and 0.7040, respectively, outperforming other deep learning models, including DeepLabv3. Compared with traditional machine learning methods, the User's Accuracy (UA), F1 score, and mIoU were improved by an average of 39.35%, 27.14%, and 39.71%, respectively. Comprehensive evaluation indicated that ENVINet5-SME exhibited superior performance in reducing misclassification and preserving the linear structure of shelterbelts. The contributions of different features varied under different site conditions, with LBP enhancing the representation of the linear structure of farmland shelterbelts and NDVI improving classification stability in areas with complex terrain and heterogeneous vegetation. Transfer learning experiments further demonstrated that the proposed method maintained stable classification accuracy under different site conditions while reducing training time by 30.56%–31.25% compared with training from scratch, indicating good generalization capability and computational efficiency. Conclusion: The proposed synergistic shelterbelt classification method (ENVINet5-SME), which integrates ENVINet5 and morphological enhancement, can effectively improve the spatial recognition accuracy, structural expression integrity, and strip target discrimination capability of farmland shelterbelts in high-resolution remote sensing images. This method demonstrates good cross-regional transferability and engineering applicability, and holds significant application value in the automatic monitoring and parameter inversion of shelterbelts at the regional scale.

Key words: farmland shelterbelts, deep learning, morphological enhancement, local binary pattern (LBP), jilin-1 satellite data