STUDY OF LAPAN SADEWA SATELLITE DATA AS A PREDICTOR OF KATULAMPA HOURLY WATER LEVEL DATA USING MACHINE LEARNING

• Main Author: Raditya Valerian, Jonathan
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/58062
• Institution: Institut Teknologi Bandung

Satellite remote sensing data is still rarely used for high accuracy spatial-temporal water level prediction. However, this study proves that satellite data from LAPAN Sadewa (sadewa.sains.lapan.go.id) with a spatial resolution of 0.01*0.01 decimal degree (30.8????????2/pixel) is able to predict the hourly water level (1 hour temporal resolution) in the Katulampa Weir (poskobanjirdsda.jakarta.go.id) with ????2?0.85 for the next 12 hours, ????2?0.8 for the next 24 hours, and missprediction ?30%. This study was conducted by conducting a sensitivity analysis on 8 data from LAPAN Sadewa (sst, qvapor, psf, rain, cloud, wind, winu, wn10), then perform spatial optimization (input extent size), temporal optimization (number of recurrent data used), and optimizing high water level reproduction. In spatial optimization, extents that are small (4px*4px) that focus on watershed length/width give the best performance (compared to 8px*8px, 16px*18px, and 28px*28px). Temporarily the optimal number of recurrent data is 24 (from t-24 hours to t-0 hours); the use of less than 24 recurrent data (3, 6, 9, 12, and 18) was still unable to achieve the performance target (R^2?0.8), and the use of more than 24 (36 and 48) did not provide a significant increase in performance. To optimize the reproduction of high discharge, the flagging algorithm (only simulating a discharge that is above the average) effectively reduces the percentage of missprediction from 57.29% (unflagged) to 19.23% (flagged). Data sensitivity analysis was performed using Deep Neural Network (2 hidden layers), and optimization was performed using Long Short Term Memory Recurrent Neural Network (LSTM RNN) with 1 LSTM layer and 2 dense layers. This study can be used to improve the existing Jakarta early warning system in Katulampa, by extending the lead-time to the next 12-24 hours; as filler for missing water level data; or various other limitless development prospects.