Exploring the Sensitivity of Recurrent Neural Network Models for Forecasting Land Cover Change

Peer reviewed: 
Yes, item is peer reviewed.
Scholarly level: 
Graduate student (PhD)
Final version published as: 

van Duynhoven, A., & Dragićević, S. (2021). Exploring the Sensitivity of Recurrent Neural Network Models for Forecasting Land Cover Change. Land, 10(3). https://doi.org/10.3390/land10030282.

Date created: 
DOI: 10.3390/land10030282
Sensitivity analysis
Recurrent neural networks
Long short-term memory
Deep learning
Land cover change modelling

Recurrent Neural Networks (RNNs), including Long Short-Term Memory (LSTM) architectures, have obtained successful outcomes in timeseries analysis tasks. While RNNs demonstrated favourable performance for Land Cover (LC) change analyses, few studies have explored or quantified the geospatial data characteristics required to utilize this method. Likewise, many studies utilize overall measures of accuracy rather than metrics accounting for the slow or sparse changes of LC that are typically observed. Therefore, the main objective of this study is to evaluate the performance of LSTM models for forecasting LC changes by conducting a sensitivity analysis involving hypothetical and real-world datasets. The intent of this assessment is to explore the implications of varying temporal resolutions and LC classes. Additionally, changing these input data characteristics impacts the number of timesteps and LC change rates provided to the respective models. Kappa variants are selected to explore the capacity of LSTM models for forecasting transitions or persistence of LC. Results demonstrate the adverse effects of coarser temporal resolutions and high LC class cardinality on method performance, despite method optimization techniques applied. This study suggests various characteristics of geospatial datasets that should be present before considering LSTM methods for LC change forecasting.

Document type: 
Natural Sciences and Engineering Research Council of Canada (NSERC)