Retrieving Water Depth Of Transparent Tufa Lakes Using UAV Optical Remote Sensing

The acquisition of high-accuracy water depth and underwater terrain is of great significance for the management and protection of tufa lake landscapes.The traditional contact bathymetry technologies(e.g.,sounding rod,sounding hammer,single/multi-beam echosounder system,etc.)not only consume a lot of manpower and material resources,but also have an impact on the aquatic environment of the lakes;airborne laser bathymetry and satellite derived bathymetry are restricted by equipment cost and spatial resolution respectively,so they are not suitable for high-frequency and high-accuracy bathymetry and underwater topographic survey of small lakes.In recent years,unmanned aerial vehicle(UAV)technology has developed rapidly and become a new remote platform,which is gradually applied to the water depth and underwater topographic mapping of clear and shallow waters such as rivers and shallow seas.In particular,the maturity and popularity of the structure-from-motion(Sf M)photogrammetry algorithms make it possible to quickly update the underwater terrain using the optical images obtained by the consumer cameras carried on the UAVs.However,compared with the general inland waters and shallow seas,tufa lakes usually have high transparency and unique spectral characteristics,resulting in some uncertainties in the process of bathymetric surveying.On the one hand,the optical detection depth of tufa lakes far exceeds the applicable depth of photogrammetry in general inland shallow waters(less than 2 m);on the other hand,there is a typical Rayleigh Scattering phenomenon in the clear tufa lakes,which affects the attenuation of blue-green light in the water.Based on the above background,this paper uses UAV images of Spark Lake obtained before the Jiuzhaigou Earthquake to produce digital elevation model and orthophoto,and constructs the bathymetric models based on them.Finally,the accuracy is verified by the dried up lake basin after the earthquake.This research content mainly includes three parts:(1)Water depth refraction correction method based on geometric analysis.The geometric analysis method is based on the digital elevation model derived from UAV photogrammetry.The initial water depth extracted from the digital elevation model is corrected through the simplified Snell’s Law(refraction correction model based on a constant factor).The results show that the corrected water depth has good overall accuracy(R~2=0.88,RMSE=1.32 m),the maximum effective depth is 12 m,but in low texture areas,it will be invalid due to the lack of effective image feature matching points.(2)Water depth inversion method based on spectral analysis.The spectral analysis method is based on the digital number(DN)values of red,green and blue bands in the orthophoto,and establishes the water depth inversion model through the selection and combination of these bands,including single band model,band ratio model and band difference model proposed in this paper.The results show that the accuracy of band difference model is significantly better than that of single band model and band ratio model,with R~2=0.86,RMSE=1.37 m,and the serious underestimation of water depth caused by shadow is improved to a certain extent.(3)Water depth measurement method based on geometric and spectral fusion with neural networks.The combined bathymetry method takes the effective water depth value after refraction correction as the reference and the band DN values in the orthophoto as the input.In the absence of measured training samples,the regression models are established through single hidden layer and multi hidden layer neural networks,and the water depths of the whole test area are predicted.The results show that the overall accuracy of the shallow neural network model with single hidden layer is the highest(R~2=0.91,RMSE=1.12 m),but the local details of water depth of the deep neural network model are better than that of the shallow neural network model,and there are no obvious local anomalies in the water depth distribution maps simulated by the two models.To sum up,this paper provides three bathymetric measurement methods of transparent tufa lakes based on UAV optical images.Among them,the method based on geometric analysis can not effectively cover the whole water area due to the single local texture,and the method based on spectral analysis can not achieve completely non-contact measurement due to the dependence on measured training samples.The fusion of geometric analysis method and spectral analysis method using neural network models realizes high-accuracy,fully-covered and non-contact bathymetry to a certain extent.At the same time,this method can be well transplanted to the bathymetry of similar lakes,and performs well in the bathymetric estimation in relatively shallow lakes.