Depth estimation of cavities from microgravity data using a new approach: the local linear model tree (LOLIMOT)

Abstract

In this paper an attempt is made to estimate depth and shape parameters of subsurface cavities from microgravity data through a new soft computing approach: the locally linear model tree, known as the LOLIMOT algorithm. This method is based on locally linear neuro-fuzzy modelling, which has recently played a successful role in various applications over non-linear system identification. A multiple-LOLIMOT neuro-fuzzy model was trained separately for each of the three most common shapes of subsurface cavities: sphere, vertical cylinder and horizontal cylinder. The method was then tested for each of the cavity shapes with synthetic data. The model’s suitability for application to real cases was analysed by adding random Gaussian noise to the data to simulate several levels of uncertainty and the results of LOLIMOT were compared to both multi-layer perceptron neural network and leastsquares minimization methods. The results showed that the LOLIMOT algorithm is more robust to noise and is also more precise than either the multi-layer perceptron or least-squares minimization method. Furthermore, the method was tested with microgravity data over a selected site located in a major container terminal at Freeport, Grand Bahamas, to estimate cavity depth and was compared to the results achieved by least-squares minimization and multi-layer perceptron methods. The proposed method can estimate cavity parameters more accurately than the least-squares minimization and multi-layer perceptron methods.