Electronic Theses and Dissertation

Data hilang merupakan permasalahan umum pada data meteorologi yang dapat
memengaruhi keakuratan analisis runtun waktu dan identifikasi tren iklim. Penelitian
ini bertujuan membandingkan kinerja beberapa metode imputasi berbasis machine
learning dalam mengisi data meteorologi yang hilang serta menganalisis tren variabel
iklim setelah imputasi. Data yang digunakan berupa data meteorologi harian periode
2016 - 2025 dari tiga wilayah yang mewakili zona waktu Indonesia, yaitu Kabupaten
Aceh Utara (WIB), Kota Samarinda (WITA), dan Kota Jayapura (WIT). Variabel yang
diamati meliputi temperatur, kelembapan udara, curah hujan, lamanya penyinaran
matahari, dan kecepatan angin. Pada tahap awal, nilai kosong pada data asli diisi
menggunakan seasonal imputation berbasis informasi day of year (DOY). Selanjutnya
dilakukan simulasi penghilangan data sebesar 15% pada variabel target untuk
mengevaluasi kinerja metode imputasi. Proses imputasi dilakukan menggunakan
Random Forest Regression (RFR), AdaBoost Regression, Support Vector Regression
(SVR), K-Nearest Neighbor Regression (KNNR), dan Kernel Ridge Regression (KRR).
Kinerja metode dievaluasi menggunakan Root Mean Squared Error (RMSE) dan
Nash-Sutcliffe Efficiency (NSE). Hasil evaluasi menunjukkan bahwa metode KRR dan
SVR memberikan kinerja terbaik dalam mengestimasi data yang hilang dibandingkan
metode lainnya. Data hasil imputasi dari metode terbaik kemudian digunakan untuk
analisis tren menggunakan uji Mann-Kendall dan estimasi Sen’s Slope. Hasil analisis
menunjukkan adanya perubahan tren pada beberapa variabel iklim di wilayah
penelitian. Secara umum, metode imputasi berbasis machine learning dapat menjadi
alternatif dalam menangani data meteorologi yang tidak lengkap sehingga analisis tren
iklim yang dihasilkan menjadi lebih representatif.

Missing data are a common issue in meteorological datasets and can affect the accuracy of time series analysis and climate trend identification. This study aims to compare the performance of several machine learning-based imputation methods for handling missing meteorological data and to analyze the trends of climate variables after the imputation process. The data used consist of daily meteorological observations from 2016 to 2025 collected from three regions representing the time zones in Indonesia, namely North Aceh Regency (WIB), Samarinda City (WITA), and Jayapura City (WIT). The observed variables include temperature, humidity, rainfall, sunshine duration, and wind speed. In the initial stage, the original missing values were first filled using a seasonal imputation approach based on day-of-year (DOY) information. Subsequently, a simulation of missing data was conducted by removing 15% of the values from the target variable to evaluate the performance of the imputation methods. The imputation process was then performed using Random Forest Regression (RFR), AdaBoost Regression, Support Vector Regression (SVR), K-Nearest Neighbor Regression (KNNR), and Kernel Ridge Regression (KRR). The performance of each method was evaluated using Root Mean Squared Error (RMSE) and Nash Sutcliffe Efficiency (NSE). The results show that KRR and SVR provide better performance than the other methods in estimating missing values. The imputed data from the best-performing methods were then used for trend analysis using the Mann Kendall test and Sen’s Slope estimator. The analysis indicates the presence of trend changes in several climate variables across the study areas. Overall, machine learning-based imputation methods can serve as an alternative approach for handling incomplete meteorological data, thereby supporting more representative climate trend analysis.