Research on Inversion Algorithms for WQP of Plateau Lakes Based on BP Neural Networks

With the improvement of computer processing speed, machine learning has developed rapidly. Combining machine learning with water quality parameters (WQP) inversion based on remote sensing data enables more extensive monitoring of water bodies. In China, the application of machine learning to remote sensing WQP monitoring started relatively late. Currently, the monitoring algorithms for four types of WQP-pH, DO, COD, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{NH}_{3}-\mathrm{H}$</tex> content-are not sufficiently mature. This paper conducts research on monitoring algorithms for these four WQP based on remote-sensing images and proposes a BP neural network algorithm suitable for monitoring and inverting WQP. Utilizing the strong nonlinear mapping capability and self-learning and self-adaptive abilities of the BP neural network (BPNN), an inversion model based on the spectral values of multiple bands from remote-sensing images is established. Experimental results indicate that the relative root mean square errors of the predicted values from the BPNN model compared to the measured values for pH, DO, COD, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{NH}_{3}-\mathrm{H}$</tex> content are 0.0808, 0.3417, 0.1970, and 0.3242, respectively. The relative errors for pH and COD test data are relatively small, and the model can effectively invert the WQP. Although the relative errors for DO and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{NH}_{3}-\mathrm{H}$</tex> content test data are larger, the goodness of fit for the inversion models of these two WQP is acceptable and meeting the basic requirements.