College of Pratacultural Science, Gansu Agricultural Universit/Key Laboratory of Grassland Ecosystems of the Mini-stry of Education/Engineering and Technology Research Center for Alpine Rodent Pest Control, National Forestry and Grassland Administration, Lanzhou 730070, China.
Sichuan Academy of Grassland Science/Qinghai-Tibet Plateau Alpine Grassland Ecology Restoration Engineering Technology Research Center/Seda Grassland Ecology Sichuan Field Scientific Observation and Research Station, Chengdu 611730, China.
Ying Yong Sheng Tai Xue Bao. 2024 Jul 18;35(7):1951-1958. doi: 10.13287/j.1001-9332.202407.020.
Rodent-infested bald spots are crucial indicators of rodent infestation in grasslands. Leveraging Unmanned Aerial Vehicle (UAV) remote sensing technology for discerning detrimental bald spots among plateau pikas has significant implications for assessing associated ecological hazards. Based on UAV-visible light imagery, we classified and recognized the characteristics of plateau pika habitats with five supervised classification algorithms, minimum distance classification (MinD), maximum likelihood classification (ML), support vector machine classification (SVM), Mahalanobis distance classification (MD), and neural network classification (NN) . The accuracy of the five methods was evaluated using a confusion matrix. Results showed that NN and SVM exhibited superior performance than other methods in identifying and classifying features indicative of plateau pika habitats. The mapping accuracy of NN for grassland and bald spots was 98.1% and 98.5%, respectively, with corresponding user accuracy was 98.8% and 97.7%. The overall model accuracy was 98.3%, with a Kappa coefficient of 0.97, reflecting minimal misclassification and omission errors. Through practical verification, NN exhibited good stability. In conclusion, the neural network method was suitable for identifying rodent-damaged bald spots within alpine meadows.
啮齿动物肆虐的光秃斑块是草原啮齿动物出没的重要指标。利用无人机遥感技术识别高原鼠兔造成的有害光秃斑块，对评估相关生态危害具有重要意义。基于无人机可见光图像，我们使用五种监督分类算法（最小距离分类（MinD）、最大似然分类（ML）、支持向量机分类（SVM）、马氏距离分类（MD）和神经网络分类（NN））对高原鼠兔栖息地的特征进行分类和识别。使用混淆矩阵评估了五种方法的准确性。结果表明，NN 和 SVM 在识别和分类高原鼠兔栖息地特征方面的性能优于其他方法。NN 对草地和光秃斑块的制图精度分别为 98.1%和 98.5%，相应的用户精度分别为 98.8%和 97.7%。总体模型精度为 98.3%，Kappa 系数为 0.97，表明分类错误和遗漏错误最小。通过实际验证，NN 表现出良好的稳定性。总之，神经网络方法适用于识别高山草甸中的啮齿动物损害光秃斑块。
