Patterns & Trends in Environmental Data / Computational Movement Analysis Geo 880

Animal movement has always been a complex field of study as it consists of and compiles the difficulties of ecology, geography and data science. This project seeks to merge these three disciplines as it tries to predict the future movement of Roe Deer through space with the usage of an aritificical neural network with constraints set up by typical biological behaviours of the Roe Deer.

• What parameters are crucial for modeling and predicting deer movement patterns using an artificial neural network (ANN), and how much does the deers gender influence these models?

• Does the articifical neural network differentiate significantly between the deers movement in different kind of landscape spaces, for example forest vs on an open grass field.

Based on the data and available data related to vegetation, a passable model that can moderate the next moving points with statistically reasonable accuracy is expected. Depending on how well we can include the the actual real world landscape into the model, the predicted movement might result in unrealistic movements (e.g. movement through water bodies or through a town).

For this project we use the data set with the movement data of the deer which was provided. In addition, we will use data with vegetation information and/or other landuse, either geodata from the canton of Zurich or the raster values from the Swiss "Arealstatistik". Which data will be used must first be compared with the literature on deer to know which vegetation / type of landscape has an influence on deer and how strong.

With our approach the amount of analytical concepts connected to movement will be rather small as the idea of the artificial neural network is that it finds those movement patterns and models of trajectories itself. It will be interesting to then analyse the results of the predictions depending on the parameters to see if there are any to us known concepts of movement or trajectory modelling.

For this paper, the main focus will be set on using R, especially the "caret"- packege which includes data training functions, including methods for implying ANN's. Futhermore the 10- fold cross validation will be used for an optimal model selection, depending on the different parameters used in the algorithm.

The biggest problem in the approach of this paper will be the parameter selection and evaluation. As movement prediction is of great difficulty, it will be hard to determine how strong a parameter should be weighted and integrated in the model. The plan B, is to alter between different algorithms, such as the k-nearest neighbors - random forest or decision tree algorithm. Additionaly the involvement of vegetation parameters can be altered.

• Are there any more variables you can think of other than the landscape type that impacts the deers movement which could be implemented into the ANN?