基于贝叶斯卷积网络的农村黑臭水体遥感识别算法研究

A remote sensing recognition algorithm for rural black and odorous water bodies on Bayesian convolutional networks

准确、快速监测农村黑臭水体分布情况是农村黑臭水体治理的重要任务.针对农村黑臭水体底数不清、人工排查效率低等问题,本文通过分析黑臭水体实测水质数据及光谱特征,选择水体清洁指数(WCI)、归一化黑臭水体指数(NDBWI)和反射率光谱指数(BOI)作为黑臭水体特征指标,基于深度学习算法,构建联合光谱与黑臭水体指数的贝叶斯多特征融合黑臭水体识别模型,实现了利用GF-2遥感影像对农村黑臭水体的遥感识别.结果表明:本文方法对六安市、阜阳市和宿州市4个农村区域黑臭水体和一般水体的识别,总体精度为86.72%,综合评价指标为80.21%,交并比为67.82%,Kappa系数为0.84,与典型阈值方法和机器学习常用方法相比,各指标均高于其他方法.此外,对比消融实验结果,贝叶斯卷积模块和全局注意力机制使本文提出的黑臭水体遥感识别算法性能得到显著提升.综上,本文方法对农村黑臭水体的监测和管理工作具有一定的指导意义.

The rapid and accurate monitoring of rural black and odorous water body distributions is an important task for treating these water bodies.In response to issues such as unclear rural black and odorous water body baselines and the low efficiency of manual inspections,this study analyzes the actual water quality data and spectral characteristics of black and odorous water bodies.It selects the water cleanliness index(WCI),normal black and odorous water body index(NDBWI),and reflectance spectral index(BOI)as characteristic indicators of black and odorous water bodies.Based on deep learning algorithms,a Bayesian multifeature fusion-centric black and odorous water body identification model is constructed;the model combines spectral and black and odorous water body indices,enabling the remote sensing-based identification of rural black and odorous water bodies using GF-2 satellite images.Experimental results indicated that the method proposed in this study achieved an overall accuracy of 86.72%,a comprehensive evaluation index of 80.21%,an intersection-over-union ratio of 67.82%,and a kappa coefficient of 0.84 in terms of identifying black and odorous water bodies and regular water bodies in three rural areas,namely,Lu′an,Fuyang,and Suzhou cities.When compared to typical thresholding methods and the commonly used machine learning techniques,this method exceeded the other approaches in terms of all evaluation metrics.Additionally,comparative ablation experiments showed that the Bayesian convolutional module and global attention mechanism significantly enhanced the performance of the remote sensing-based black and odorous water body identification algorithm proposed in this study.In conclusion,this method provides valuable guidance for the monitoring and management of rural black and odorous water bodies.