摘要
受青藏高原气候暖湿化影响,青海湖新生湖滨带刚毛藻水华频繁暴发。以往刚毛藻水华提取研究主要依赖多源卫星遥感影像,但受限于影像空间分辨率和混合像元效应,难以精确捕捉刚毛藻水华的真实分布及其细节特征。本文利用低空无人机影像结合Attention DeepLab V3+深度学习模型自动提取青海湖刚毛藻水华特征,对比分析其与光谱指数和机器学习方法的提取结果,并探讨无人机影像与光学卫星遥感影像提取结果的差异。结果表明:(1)Attention DeepLab V3+可在没有先验阈值情况下准确检测刚毛藻水华分布范围,模型的Kappa系数、精度、召回率和F1得分分别为0.985、0.969、0.983和0.976,表明识别能力较强。(2)与随机森林模型和红-绿-蓝浮游藻类指数相比,该模型Kappa系数和F1得分分别提高4.47%~29.75%和6.35%~34.02%,能够更好地适应复杂的刚毛藻水华分布特征,尤其是在边界细节呈现和空洞分离方面具有明显优势。(3)基于Landsat OLI-2和Sentinel-2 MSI等常用光学卫星遥感影像的提取结果存在高估青海湖刚毛藻水华面积的现象,前者平均相对误差值范围为65.28%~110.69%,后者平均相对误差值范围为5.5%~323.47%。本研究利用无人机影像的高分辨率优势,为准确评估青海湖刚毛藻水华的真实分布提供了技术支持,并为其他水体藻华特征的监测与追踪奠定了基础。
Frequent outbreaks of Cladophora blooms in the newly formed littoral zone of Lake Qinghai have been observed due to the warming and humidification of the Qinghai-Tibet Plateau climate.Previous studies on the extraction of Cladophora blooms mainly relied on multi-source satellite remote sensing imagery.However,the limitations of image spatial resolution and mixed-pixel effects hindered the accurate identification of the true distribution and detailed features of the blooms.This study utilized low-altitude UAV imagery combined with the Attention DeepLab V3+deep learning model to automatically extract Cladophora bloom features in Lake Qinghai.A comparative analysis was conducted with results derived from spectral indices and machine learning methods,and the differences between UAV imagery and optical satellite remote sensing imagery in extracting Cladophora blooms were explored.The subsequent results are outlined below:(1)It has been demonstrated that Attention DeepLab V3+is capable of accurately detecting Cladophora blooms without the necessity of prior thresholds,achieving a kappa coefficient,precision,recall,and F1 score of 0.985,0.969,0.983,and 0.976,respectively.(2)In comparison with both the random forest model and the red-green-blue floating algae index,the model demonstrated a marked improvement in both the kappa coefficient and the F1 score,with increases of 4.47%and 6.35%,respectively.This is particularly noteworthy in terms of its superior adaptability to complex bloom distribution patterns,as evidenced by its ability to capture boundary details and differentiate between voids.(3)Optical satellite remote sensing imagery tends to overestimate Cladophora blooms in Lake Qinghai,with mean relative error values ranging from 5.5%to 323.47%.This study utilized the high-resolution capabilities of UAV imagery to provide technical support for accurately assessing the true distribution of Cladophora blooms in Lake Qinghai,thereby establishing a foundation for the monitoring and tracking of algal bloom features in other water bodies.
作者
张娟
姚晓军
陈进轩
张瑜轩
韩胜利
窦皓敏
Zhang Juan;Yao Xiaojun;Chen Jinxuan;Zhang Yuxuan;Han Shengli;Dou Haomin(College of Geography and Environmental Science,Northwest Normal University,Lanzhou 730070,P.R.China;Qinghai Lake National Nature Reserve Administration,Gangcha 812300,P.R.China;Qinghai Lake Comprehensive Observation and Research Station,Chinese Academy of Sciences,Gangcha 812300,P.R.China)
出处
《湖泊科学》
北大核心
2026年第1期129-141,I0014,I0015,共15页
Journal of Lake Sciences
基金
国家自然科学基金项目(42571163)
西北师范大学绿洲科学科研成果突破行动计划(NWNU-LZKX-202301)联合资助。
