GEOSPATIAL PREDICTIVE MODEL BASED ON DEEP LEARNING FOR FORECASTING THE FREQUENCY OF CLOUD OCCURRENCE, CASE STUDY ON LAND AREA OF EASTERN JAVA

• Main Author: Luthfi Hadiyanto, Ahmad
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71026
• Institution: Institut Teknologi Bandung

The characteristic of clouds over a region is related to the weather and climate in the area. Clouds affect energy balance by blocking solar radiation from the sun and trapping the energy emitted from the surface of the earth. The net process affects climate over an area. Clouds are also part of the hydrological cycle. Cloud is formed by lifted warm water vapor which will drain as precipitation. The energy will be released and cool the weather. The knowledge on clouds occurence over an area will also be helpful to plan the satellite data recording. Information on the prediction of cloud occurrence in the future will minimize the use of satellite resources. Therefore, this research tries to build a forecasting model to estimate percentage/probability of clouds occurrence using geospatial model. State of the art of clouds prediction applies numerical weather prediction. The model was built based on physical process of fluids with numerical computation. Because of complexity on atmospheric system, the computation in numerical weather prediction needs high resources, especially for long-term forecasting. The use of geospatial model with remote sensing satellite data and deep learning method may reduce computation time and complexity of the model in this research. Previous research on clouds forecasting over land using remote sensing satellite data resulted in short term prediction. The model was built with limited usage of satellite data and was performed on specific conditions. The model observes clouds and predicts the future state of the clouds. In this research, forecasting model is built using temporal data of surface parameters and low atmosphere parameters to achieve long term prediction. Aggregation and interpolation are performed to Himawari-8 spectral data to get full spatial resolution and continuous temporal data. This method will be applied to get a general model which can be applied in any condition of atmosphere. Hypothesis testing has been done on several spectral bands of Himawari-8 data. Dynamic parameters will be utilized by applying aggregation every three-month period. Other parameters with slower change will be processed to a one-month period of aggregate. The output of the model is percentage of cloud occurrence in the next month. The value is calculated using cloud classification data on an hourly basis. The result shows that there are two groups of variables passing hypothesis testing on different periods. Spectral data related to humidity and land cover type have high correlation coefficient to cloud percentage in February to April. Spectral data related to surface temperature have higher correlation coefficient to cloud percentage in May to July. Therefor there are two models built and each of them has a different set of independent variables. The model is built for land area over eastern Java using Convolutional Neural Network (CNN). The first model results in good accuracy with Normalized RMSE (NRMSE) 2.191% to predict cloud percentage in February, March, and April 2020. The second model has lower accuracy with NRMSE 7.309% to predict cloud percentage in May, June, and July 2020. Compared to land cover type, this research shows that error prediction is high over some places over forest. It raises suggestion to separate prediction model into two categories: forest, and non-forest. The benefit of the model can also be optimized by selecting a specific type of cloud in a specific time span.