Soil Moisture Prediction Model on Peatlands using Long Short-Term Memory

Abstract

Peatlands play an important role in maintaining global ecosystem balance, but are highly susceptible to fires due to decreased soil moisture. This study aims to predict soil moisture on peatlands using the Long Short-Term Memory (LSTM) algorithm as a time series learning model. The data used includes variables of soil moisture (GWETPROF), rainfall (PRECTOTCORR), and temperature (T2M) obtained from NASA Langley Research Center's Prediction of Worldwide Energy Resources (POWER) for the period from August 1, 2019, to December 31, 2023. The preprocessing involved identifying and handling missing values using the mean imputation method and normalization with the Min-Max Scaling technique. Correlation analysis showed a weak relationship between variables, so all of them were used as independent features. The LSTM model was built with parameters of 50 neurons, ReLU activation function, Adam optimizer, and a dropout rate of 0.2. The test results showed that the model was able to accurately predict water content with a MAE value of 0.005, MSE ≈ 0.000, RMSE of 0.014, and R² of 0.97 on the test data. These results indicate that LSTM is effective in capturing temporal patterns and fluctuations in soil moisture, making it a potential tool for more adaptive and data-driven peatland fire mitigation