Machine Learning Methods for Predicting Manure Management Emissions

Abstract

Indonesia is committed to reducing greenhouse gas (GHG) emissions through a nationally determined contribution (NDC) scheme. The target to reduce GHG emissions is 29% through the business as usual (BAU) scheme or 41% with international aid. These ambitious targets require transformations in energy, food, and land-use systems, which need to cope with the potential trade-offs among many targets, such as food security, energy security, avoided deforestation, biodiversity conservation, land use competition, and freshwater use. Mitigation and adaptation have complementary roles in responding to climate change at both temporal and spatial scales. This paper aims to perform simulations and predictions on manure management emissions producing CO2eq using machine learning methods of long short-term memory (LSTM) and gated recurrent unit (GRU). The hidden layer architecture used was six combinations, while the dataset was obtained from the fao.org repository. The optimizer used in this paper was RMSprop, with a graphical user interface using the Streamlit dashboard. The results of this study are (a) cattle with fifteen epochs using hidden layer four combinations (LSTM, GRU, LSTM, GRU) yielded RMSE 450,601; (b) non-dairy cattle with fifteen epochs and one hidden layer (GRU, GRU, GRU, GRU) yielding RMSE 361.421; (c) poultry birds with twelve epoch values and three hidden layers (GRU, GRU, LSTM, LSTM) resulted in an RMSE value of 341.429. The challenges faced were the determination of epochs, the combination of hidden layers, and the characteristics of the relatively small number of datasets. The results of this study are expected to provide added value for developing better decision support tools and models to assess emission trends in the livestock sector and develop CO2eq emission mitigation strategies that lead to sustainable fertilizer management practices.