Evaluating long short-term memory networks for modeling land cover change

Land cover change (LCC) can be viewed as dynamic complex systems which require relevant relationships to be encoded when represented within various modeling approaches. Recurrent Neural Networks (RNNs), specifically the Long Short-Term Memory (LSTM) variant, belong to a category of Deep Learning (DL) approaches best suited for sequential and timeseries data analysis, thus suitable for representing LCC. The primary objective of this study is to examine the capacity and effectiveness of LSTM networks for forecasting LCC given varying geospatial input datasets with feature impurities. Using synthetic and MODIS land cover datasets for British Columbia, Canada, results demonstrate the sensitivity of LSTM models to varying geospatial input dataset characteristics. Geospatial datasets with finer temporal resolutions and increased timesteps yielded favourable results while coarser temporal resolutions and fewer timesteps were affiliated with less successful outcomes. This thesis research contributes to the advancement of automated, data-driven DL methodologies for forecasting LCC.