[Paper Review] Modeling the spatio-temporal dynamics of land use change with recurrent neural networks

Guodong Du, Liang Yuan+2 more|arXiv (Cornell University)|Mar 29, 2018

Land Use and Ecosystem Services54 references3 citations

TL;DR

This study proposes using recurrent neural networks (RNNs), particularly LSTM and LSTM-peephole variants, to model and forecast spatio-temporal land use change in Tsukuba, Japan, using annual land use data from 2000 to 2016. The gated RNN architectures significantly outperform the simple RNN, with LSTM-peephole achieving the highest performance, demonstrating the importance of modeling long-term temporal dependencies in land use dynamics.

ABSTRACT

This study applies recurrent neural networks (RNNs), which are known for its ability to process sequential information, to model the spatio-temporal dynamics of land use change (LUC) and to forecast annual land use maps of the city of Tsukuba, Japan. We develop two categories of RNN models: 1) simple RNN, which is the basic RNN variant, 2) three RNN variants with advanced gated architecture: long short-term memory (LSTM), LSTM with peephole connection (LSTM-peephole), and gated recurrent unit (GRU) models. The four models are developed using spatio-temporal data with high temporal resolution, annual data for the periods 2000 to 2010, 2011 and 2012 to 2016 are used for training, validation and testing, respectively. The predictive performances are evaluated using classification metrics (accuracy and F1 score) and the map comparison metrics (Kappa simulation and fuzzy Kappa simulation). The results show that all RNN models achieve F1 scores higher than 0.55, and Kappa simulations higher than 0.47. Out of the four RNN models, LSTM and LSTM-peephole models significantly outperform the other two RNN models. Furthermore, LSTM-peephole model slightly outperforms the LSTM model. In addition, the results indicate that the RNN models with gated architecture, which have better ability to model longer temporal dependency, significantly outperform the simple RNN model. Moreover, the predictive performance of LSTM-peephole model gradually decreases with the decrease of temporal sequential length of the training set. These results demonstrate the benefit of taking temporal dependency into account to model the LUC process with RNNs.

Motivation & Objective

• To model the spatio-temporal dynamics of land use change (LUC) using sequential data over time.
• To evaluate the performance of various RNN architectures in forecasting annual land use maps.
• To assess the impact of temporal dependency modeling on predictive accuracy in LUC forecasting.
• To compare the effectiveness of simple RNNs versus gated RNNs (LSTM, LSTM-peephole, GRU) in capturing long-term land use trends.

Proposed method

• The study employs four RNN variants: simple RNN, LSTM, LSTM with peephole connections, and GRU, to process sequential land use data.
• Spatio-temporal data with annual resolution from 2000 to 2016 is used, with training (2000–2010), validation (2011), and testing (2012–2016) splits.
• Model performance is evaluated using classification metrics (accuracy, F1 score) and map comparison metrics (Kappa simulation, fuzzy Kappa simulation).
• The models are trained to predict land use transitions by learning temporal dependencies across consecutive years.
• The gated architectures (LSTM and LSTM-peephole) are designed to better capture long-term dependencies through learnable memory gates.
• The study analyzes the sensitivity of model performance to the length of the temporal sequence in the training set.

Experimental results

Research questions

• RQ1 Can RNNs effectively model and forecast spatio-temporal land use change using annual land use data?
• RQ2 Do gated RNN architectures such as LSTM and LSTM-peephole outperform simple RNNs in land use forecasting?
• RQ3 How does the length of the temporal training sequence affect the predictive performance of RNN models in LUC modeling?
• RQ4 To what extent does modeling long-term temporal dependencies improve the accuracy of land use change predictions?

Key findings

• All RNN models achieved F1 scores higher than 0.55 and Kappa simulation values above 0.47, indicating strong predictive performance.
• The LSTM and LSTM-peephole models significantly outperformed the simple RNN and GRU models in both F1 score and Kappa simulation.
• The LSTM-peephole model slightly outperformed the standard LSTM model, suggesting enhanced memory control improves forecasting.
• The predictive performance of gated RNNs, especially LSTM-peephole, decreased with shorter temporal sequences in the training set, highlighting their reliance on long-term context.
• The results confirm that modeling long-term temporal dependencies is crucial for accurate land use change forecasting using RNNs.
• The study demonstrates that gated RNNs are better suited than simple RNNs for capturing complex spatio-temporal dynamics in land use change.