Walnut tree canopy information extraction based on unmanned aerial vehicle multispectral images

【Objective】Accurate acquisition of vegetation type, nutritional status, and health condition is crucial in ecological monitoring and agricultural research. Such information not only aids in the correct identification of different land cover types but also establishes a solid foundation for in-depth studies on inversion models of walnut tree canopy parameters. With the rapid development of unmanned aerial vehicle (UAV) remote sensing technology, high- resolution multispectral imaging has become an effective means to obtain vegetation canopy information at the individual plant scale. However, due to spectral mixing between the target canopy and background features such as shadows, adjacent plants, and weeds, accurately extracting walnut canopy information remains a challenging task. This study aims to systematically evaluate the performance of six supervised classification methods in extracting walnut tree canopy information from UAV multispectral imagery, and to identify the most suitable method for precise canopy segmentation.【Methods】In this study, the Wen 185 walnut tree canopy was used as the research object. UAV multispectral images were acquired as the data source using a DJI M300 RTK drone equipped with an MS600 Pro multispectral camera, which captured six spectral bands: blue (450 nm), green (550 nm), red (660 nm), red edge 1 (720 nm), red edge 2 (750 nm), and near-infrared (840 nm). The flight was conducted at an altitude of 100 m, achieving a spatial resolution of 3.33 cm, with both heading and side overlaps set at 80%. Image preprocessing, including registration, stitching, and cropping, was performed using Pix4D mapper and ENVI 5.3. Six supervised classification methods were applied: Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Support Vector Machine (SVM), and Neural Network. Training and validation samples were defined using Region of Interest (ROI) tools, with a 7∶3 split between training and validation sets. The Jeffries-Matusita and Transformed Divergence methods were used to evaluate the separability of training samples. After initial classification, post-processing techniques including Clump, Sieve, Aggregate, and Majority/ Minority analysis were applied to optimize the classification results. Accuracy assessment was conducted using confusion matrices, overall accuracy, Kappa coefficient, commission error, omission error, producer accuracy, and user accuracy.【Results】The study revealed significant differences in the performance of the six classifiers before and after post-processing. The Support Vector Machine (SVM) classifier, when combined with Majority processing, achieved perfect performance across all evaluation metrics: an overall accuracy of 100%, a Kappa coefficient of 1.000 0, and both commission and omission errors of 0%. Additionally, the user accuracy and producer accuracy for both the walnut canopy and other land cover categories reached 100% , demonstrating the method's exceptional capability in distinguishing target vegetation from complex backgrounds. The Maximum Likelihood and Mahalanobis Distance classifiers also exhibited excellent results following aggregation or clustering optimization, each attaining an overall accuracy of 99.93% with a Kappa coefficient of 0.997 8. These methods showed strong robustness in handling spectral mixing issues, though minor misclassification persisted in nontarget categories. The Neural Network classifier, when enhanced with Majority processing, reached an overall accuracy of 99.88% (Kappa = 0.998 1); however, it presented a slight omission error of 0.54%, indicating a minimal tendency to under- predict walnut canopy pixels. In contrast, the Minimum Distance classifier, even after post-processing such as aggregation, attained only 85.97% overall accuracy, with high commission errors (41.1%) observed particularly in non-canopy classes, reflecting limited discriminative power in heterogeneous areas. The Parallelepiped classifier performed least favorably, with a final accuracy of merely 64.19% after clustering optimization, accompanied by significant omission and commission errors—26.23% for walnut and 59.78% for other features—highlighting its inadequacy in handling complex spectral variations within high- resolution imagery. These results underscored the critical influence of classifier selection and post-processing strategy on extraction accuracy. Advanced methods like SVM and Maximum Likelihood, especially when paired with appropriate spatial optimization techniques, were highly effective for precise canopy delineation, whereas traditional algorithms exhibited notable limitations in high-precision agricultural remote sensing applications.【Conclusion】The combination of Support Vector Machine (SVM) and Majority patch processing is identified as the optimal method for extracting walnut tree canopy information under the current research conditions, demonstrating extremely high classification accuracy and stability. This method achieves a perfect overall ac-curacy of 100%, a Kappa coefficient of 1.0, and zero commission and omission errors, confirming its superior capability in distinguishing walnut canopies from complex backgrounds such as shadows, weeds, and adjacent vegetation. Its robustness is further evidenced by consistent user and producer accuracies of 100% for all land cover categories, making it highly suitable for precise canopy segmentation at the individual plant scale. Other advanced classifiers, including Maximum Likelihood, Mahalanobis Distance, and Neural Networks, also show high potential when supported by appropriate post- processing techniques. For instance, Maximum Likelihood and Mahalanobis Distance reached an overall accuracy of 99.93% (Kappa=0.997 8) with aggregation or clustering, while the Neural Network approach achieved 99.88% accuracy (Kappa=0.998 1) when combined with Majority processing, despite a slight omission error of 0.54% . These methods exhibit strong performance in handling spectral mixing and maintaining spatial consistency. In contrast, traditional classification methods such as Minimum Distance and Parallelepiped proved inadequate for high-precision extraction tasks. Even after post-processing, the Minimum Distance classifier reached only 85.97% accuracy, with notably high commission errors (41.1% ) in non-target categories. The Parallelepiped classifier performed least favorably, with an accuracy of 64.19%, along with substantial omission errors (26.23% for walnut, 59.78% for others), underscoring its limitations in feature discrimination and applicability in heterogeneous agricultural environments. These results highlight the importance of selecting modern high-dimensional classifiers coupled with spatial optimization techniques for accurate vegetation mapping in precision agriculture applications.