Open AccessPrediction of Water Levels on Peatland using Deep Learning
Author(s) -
Namora,
Jan Everhard Riwurohi
Publication year - 2022
Publication title -
jurnal resti (rekayasa sistem dan teknologi informasi)
Language(s) - English
Resource type - Journals
ISSN - 2580-0760
DOI - 10.29207/resti.v6i2.3919
Subject(s) - peat , environmental science , water level , artificial neural network , hydrology (agriculture) , meteorology , machine learning , statistics , computer science , mathematics , engineering , geography , cartography , archaeology , geotechnical engineering
The water level on peatlands is one of the causes of peatland fires, so water levels must be maintained at a safe standard value. Government Regulation No. 71/2014 stipulates water level standard value is 0.4 meters. The forest and land fires in 2015 caused huge losses of 220 trillion Rupiah. However, fires still occur frequently. BRGM (Peatland and Mangrove Restoration Agency) installed sensors measuring peatland water levels to obtain real-time water level data. These data can be used to predict water levels. Several previous studies used drought indices, regression models, and artificial neural networks to predict water levels. In this study, it is proposed to use deep learning Long Short-Term Memory (LSTM), and apply the CRISP-DM methodology. The dataset in this study contains water level data from 15 measurement stations in Central Kalimantan from 2018 through 2021. It was concluded that the LSTM model was able to predict water level well, as indicated by the average RMSE of 0.07 m, the average R2 of 0.85, and the average MAE of 0.04 m. The optimal LSTM model parameters are 50 epochs, a 70%:30% ratio of training data to testing data, and 2 hidden layers.