Novel Deep Learning Approaches for Mapping Variation of Ground Level from Spirit Level Measurements

Abstract

Abstract. This study investigates the use of new machine learning techniques in mapping variation in ground levels based on ordinary spirit levelling (SL) measurements. Convolution Neural Network (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and bi-directional LSTM (BI-LSTM) were developed and compared in the current study to estimate the leveling through SL measurements. SL measurements of the Manzalla region, Egypt, were used in the current study. 3253 datasets of SL observation points, including 229 benchmarks of precise levelling (PL), were used to design and verify the proposed model’s results. The results show the developed LSTM model outperforms CNN, RNN, and BI-LSTM in modeling ground leveling in the training and testing stages. The root mean square error and correlation determination of the LSTM model are 7.4 cm and 0.99, respectively, in the testing stage. The accuracy of mapping ground levelling through the developed LSTM model is close to 99 % in terms of model error.