Sustainable Career Transition Pathway: Bridging Academia, Public Sector, and Industry - Spatiotemporal modelling for rainfall forecasting: using social media data to support sustainable career

Can social media and satellite data help predict rainfall more accurately? This study, introduces a powerful spatiotemporal model, GSTAR-NN, that combines machine learning and climate data to support both environmental forecasting and sustainable research careers.

Budi Nurani Ruchjana

Why rainfall forecasting matters?

In daily life, most processes are spatiotemporal, and the data are generated both by where (spatial) and when (temporal) observations happen. Following the 13th pillar of the Sustainable Development Goals (SDGs) of climate change, rainfall is one of the most essential processes, and rainfall forecasting has become a crucial part of our environment.

In this research, led by Budi Nurani Ruchjana and Atje Setiawan Abdullah from Padjadjaran University and Devi Munandar from the National Research and Innovation Agency, both in Indonesia, we propose spatiotemporal modelling for rainfall forecasting. We use social media data, especially on climate phenomena, as big data provided by the National Aeronautics and Space Administration Prediction Of Worldwide Energy Resources (NASA POWER). We propose a spatiotemporal modelling combined with a machine learning approach based on time series analysis called the Generalized Space- Time Autoregressive Neural Network (GSTAR- NN model). How the model works

Bionote

Budi Nurani Ruchjana received her BSc in Mathematics from Padjadjaran University in Indonesia in 1987, an MSc in Applied Statistics from Institut Pertanian Bogor in Indonesia in 1992, and PhD in Mathematics and Natural Sciences from Institut Teknologi Bandung in Indonesia in 2002. Currently, she is a full professor at the Department of Mathematics at Padjadjaran University. Her research interests include spatiotemporal modelling, stochastic processes, time series analysis, spatial analysis, geostatistics, mathematics and science data, and ethnomathematics.

How the model works

The GSTAR model is a special case of Vector Autoregressive (VAR) as a multivariate time series model, and Neural Network (NN) as part of Machine Learning, which has a powerful model for studying complex and abstract data features, especially in non- linear data. The integration of GSTAR-NN is divided into two steps. First, the model is used to calculate the residuals of GSTAR, the errors between the prediction and actual values, and second, we use NN as a feature extractor for the residual input from the GSTAR results. It can be based on changing the non-linear data pattern into a higher numerical representation through the layers in the Neural Network.

We applied this combined GSTAR-NN model to big data from NASA POWER for climate phenomena in Indonesia using the data analytics lifecycle methodology. We followed six stages of research:

• Discovery. Identifying problems;
• Data identification. Determining data sources and hypotheses;
• Data preparation. Cleaning, transforming, and data storage;
• Model planning. For GSTAR, NN, and their integration process;
• Model building. Training, testing, and model validation;
• Communication. Evaluation of the model and final stage of discussing implementation with practitioners.

In the development of the GSTAR-NN model, we assume that:

• The data is stationary, meaning it doesn’t change over time or space.
• The GSTAR model assumes a linear relationship between dependent and independent variables.
• The homogeneity of error variance across the dataset.
• The weight matrix is based on inverse distance.
• The Ordinary Least Squares method is used to estimate parameters.

Our case study in developing of the GSTAR- NN model is focused on rainfall data for three locations in the West Java region in Indonesia: Lembang, Bogor, and Sukabumi. We used a model structure GSTAR(1,1) which means that the rainfall at i location at time t, is influenced by the rainfall data lagged by a one-time unit and the rainfall in the surrounding locations, and an error term.

For the weight initialisation method with the ReLU function, the variance of the weights is calculated based on the number of inputs in the layer, as in the He method. However, there is a change to the calculation of the variance factor in the formula. We applied the GSTAR- NN(1,1) architecture with three locations: one hidden layer, six input neurons, q neurons in the hidden layer, and one neuron in the output layer.

Real impact and collaboration

The results showed that spatiotemporal prediction using the GSTAR(1,1)-NN model reduced computational costs while improving accuracy which is measured using Mean Absolute Percentage Error. Visual reports from the model can help government agencies and organisations make better-informed decisions about rainfall forecasting.

Furthermore, the study can be developed through research collaboration between academics and practitioners in the field of climate change. We also created an open- source tool in R, so we can work together with an expert in the computer industry to make the programme more user-friendly and easier for society to use.

Finally, we implemented the result in the
international consortium of Research
Innovation and Staff Exchange Social Media
Analytics (RISE_SMA) funded by the European
Union during 2019–2024. It is coordinated
by Professor Stefan Stieglitz from Potsdam
University and involves different partners,
including Leiden University, Agder University,
Queensland University, Sydney University,
UNISINOS University, Padjadjaran University,
VIRTIMO Berlin and Kristiansand Municipality.
All members contributed significantly to the
collaboration, helping to advance research,
support sustainable careers, and create
practical tools that benefit both society and
industry.

Budi Nurani Ruchjana
Orcid
X
Padjadjaran University
budi.nurani@unpad.ac.id