Water level dynamics in catchment-scale rivers is an important factor for surface water studies. Manual measurement is highly accurate but inefficient. Using automatic water level sensors has disadvantages such as high cost and difficult maintenance. In this study, a water level recognition method based on digital image processing technology and CNN is proposed. For achieving batch segmentation of source images, the coordinates of the water ruler region in the source image and characters’ region and the scale lines’ region on the ruler are obtained by using image processing algorithms such as grayscale processing, edge detection, and the tilt correction method based on Hough-transform and morphological operations. The CNN is then used to identify the value of digital characters. Finally, the water level value is calculated according to the mathematical relationship between the number of scale lines detected by pixel traversal in the binarized image and the value of digital characters. This method is used to identify the water levels of the water ruler images collected in the Hulu watershed of the Qilian Mountains in Northwest China. The results show that the accuracy compared with the actual measured water level reached 94.6% and improved nearly 24% compared to the template matching algorithm. With high accuracy, low cost, and easy deployment and maintenance, this method can be applied to water level monitoring in mountainous rivers, providing an effective tool for watershed hydrology research and water resources management.
The fluxes of sensible heat (H) and latent heat (LE), which are generally the important parts of the energy and mass balances over glacier surfaces, are widely quantified by the bulk method. However, due to the difficulty of determining the aerodynamic roughness length z0m in this method, H and LE values may still have large uncertainties with significant inaccuracy. To acquire reliable varying and intrinsic z0m values, new simpler parameterizations for z0m values at different ranges of the friction velocity u*b were fitted in this study. The method was implemented using the related meteorological data and glacial sublimation/condensation measured at the top of the August‐one glacier (4817 m asl) in the Qilian Mountains from 1 July 2016 to 15 August 2017. The parameterization shows that z0m increased sharply when u*b exceeded 0.43 m/s (the approximate threshold value) and that the effect of snowdrift was slight for the hourly z0m values in the range 0.15 ≤ u*b ≤ 0.43 m/s, which could thus be used to calculate the daily z0m. During the wet period (1 July to 24 September 2016 and 5 May to 15 August 2017), the turbulent fluxes calculated by the bulk method showed that net radiation Rnet was the primary source of surface energy (60.7 w/m2), whereas during the dry period (25 September 2016 to 4 May 2017), the main heat sink was the positive H (28.5 w/m2) rather than the negative Rnet (−10.0 w/m2).
As an essential data-driven model, machine learning can simulate runoff based on meteorological data at the watershed level. It has been widely used in the simulation of hydrological runoff. Considering the impact of snow cover on runoff in high-altitude mountainous areas, based on remote sensing data and atmospheric reanalysis data, in this paper we established a runoff simulation model with a random forest model and ANN (artificial neural network) model for the Xiying River Basin in the western Qilian region The verification of the measured data showed that the NSE (Nash–Sutcliffe efficiency), RMSE (root mean square error), and PBIAS (percent bias) values of the random forest model and ANN model were 0.701 and 0.748, 6.228 m3/s and 4.554 m3/s, and 4.903% and 8.329%, respectively. Considering the influence of ice and snow on runoff, the simulation accuracy of both the random forest model and ANN model was improved during the period of significant decreases in the annual snow and ice water equivalent in the Xiying River Basin from April to May, after the snow remote sensing data were introduced into the model. Specifically, for the random forest model, the NSE increased by 0.099, the RMSE decreased by 0.369 m3/s, and the PBIAS decreased by 1.689%. For the ANN model, the NSE increased by 0.207, the RMSE decreased by 0.700 m3/s, and the PBIAS decreased by 1.103%. In this study, based on remote sensing data and atmospheric reanalysis data, the random forest model and ANN model were used to effectively simulate hydrological runoff processes in high-altitude mountainous areas without observational data. In particular, the accuracy of the machine learning simulations of snowmelt runoff (especially during the snowmelt period) was effectively improved by introducing the snow remote sensing data, which can provide a methodological reference for the simulation and prediction of snowmelt runoff in alpine mountains.