Optimizing Neural Network for Parameter Estimation of Highly Multivariate Log Gaussian Cox Process Using Dropout Training

Analyzing highly multivariate spatio-temporal point pattern data is very challenging, especially using the standard procedure since it cannot handle huge data volume, complex spatio-temporal model, and expensive computation. Meanwhile, neural networks have shown their ability to handle complex problems. This study uses a robust neural network model with dropout layers to estimate parameters of highly multivariate spatio-temporal log Gaussian Cox processes. We employ our model to assess the distributional patterns of 25 tree species within Barro Colorado Island dataset, observed at 4 different timestamps. We achieved an accuracy improvement of more than 2.5% over previous state-of-the-art work, demonstrating that our network is better to handle highly multivariate spatio-temporal data.