Update Artikelilmiah: 54855

DIAN RIZQI SAPUTRA

NIM

Judul Artikel

Abstrak (Bhs. Indonesia)

Wilayah Cilacap merupakan daerah pesisir tropis dengan dinamika atmosfer yang kompleks sehingga memerlukan informasi prediksi cuaca yang akurat untuk mendukung berbagai sektor industri dan mitigasi bencana hidrometeorologi. Penelitian ini bertujuan menganalisis kontribusi indeks stabilitas atmosfer hasil radiosonde terhadap prediksi kondisi cuaca serta menginterpretasikannya berdasarkan karakteristik fisis atmosfer, mengevaluasi kinerja algoritma Extreme Gradient Boosting (XGBoost), Random Forest (RF), dan Support Vector Machine (SVM) pada berbagai tingkat kompleksitas kategori, kondisi musim, dan jumlah indeks, serta mengevaluasi kesesuaian hasil prediksi terhadap kondisi cuaca yang teramati dan membandingkannya dengan prediksi BMKG berbasis Numerical Weather Prediction (NWP). Data yang digunakan berupa observasi radiosonde dan laporan sinoptik BMKG Cilacap periode 2015–2024 dengan enam indeks stabilitas atmosfer, yaitu Showalter Index (SI), Lifted Index (LI), K Index (KI), Total Totals Index (TTI), Severe Weather Threat Index (SWEAT), dan Convective Available Potential Energy (CAPE). Analisis karakteristik indeks dilakukan menggunakan matriks korelasi, Principal Component Analysis (PCA), dan Random Forest Feature Importance, sedangkan evaluasi model dilakukan pada skema klasifikasi dua, tiga, dan empat kategori, variasi musim, serta percobaan reduksi jumlah indeks. Hasil analisis menunjukkan bahwa KI, CAPE, dan SWEAT merupakan indeks yang memberikan kontribusi dominan terhadap proses klasifikasi cuaca. Peningkatan kompleksitas klasifikasi menurunkan kinerja model akibat meningkatnya kemiripan karakteristik atmosfer antar kategori, sedangkan kemampuan model dalam mendeteksi hujan lebih tinggi pada musim hujan dibandingkan musim kemarau. Percobaan reduksi menunjukkan bahwa penggunaan empat indeks, yaitu KI, CAPE, SWEAT, dan TTI, mampu mempertahankan kinerja model yang relatif serupa dengan penggunaan enam indeks. Di antara algoritma yang diuji, SVM menunjukkan kinerja yang paling konsisten pada berbagai skenario pengujian. Evaluasi terhadap laporan sinoptik dan citra satelit Himawari menunjukkan bahwa hasil prediksi lebih konsisten pada kondisi atmosfer dengan karakteristik yang jelas, sedangkan ketidaksesuaian lebih sering terjadi pada kondisi atmosfer transisi. Perbandingan pada periode pengujian tahun 2024 menunjukkan bahwa SVM memiliki kemampuan deteksi hujan yang relatif sebanding dengan prediksi BMKG berbasis NWP, meskipun NWP memiliki kemampuan yang lebih seimbang dalam membedakan kondisi cerah dan hujan. Hasil penelitian menunjukkan bahwa machine learning berbasis indeks stabilitas atmosfer dapat dimanfaatkan sebagai pendekatan pendukung dalam prediksi kondisi cuaca di wilayah tropis Cilacap.

Abtrak (Bhs. Inggris)

Cilacap is a tropical coastal region characterized by complex atmospheric dynamics, requiring accurate weather prediction information to support various industrial sectors and hydrometeorological disaster mitigation. This study aims to analyze the contribution of radiosonde-derived atmospheric stability indices to weather condition prediction and interpret their contributions based on the physical characteristics of the atmosphere; evaluate the performance of Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machine (SVM) algorithms across different levels of classification complexity, seasonal conditions, and numbers of indices; and evaluate the consistency of the prediction results with observed weather conditions and compare their performance with BMKG weather forecasts based on Numerical Weather Prediction (NWP). The dataset consisted of radiosonde observations and SYNOP reports from BMKG Cilacap for the 2015–2024 period, using six atmospheric stability indices: Showalter Index (SI), Lifted Index (LI), K Index (KI), Total Totals Index (TTI), Severe Weather Threat Index (SWEAT), and Convective Available Potential Energy (CAPE). Atmospheric stability characteristics were analyzed using a correlation matrix, Principal Component Analysis (PCA), and Random Forest Feature Importance, while model performance was evaluated using two-, three-, and four-category classification schemes, seasonal variations, and a feature reduction experiment. The results of the study show that KI, CAPE, and SWEAT provide dominant contributions to weather classification. Increasing classification complexity reduced model performance due to greater overlap in atmospheric characteristics among categories, whereas the ability to detect rainfall was higher during the rainy season than during the dry season. The feature reduction experiment showed that using four indices, namely KI, CAPE, SWEAT, and TTI, was able to maintain model performance at a level relatively similar to that obtained using all six indices. Among the evaluated algorithms, SVM demonstrated the most consistent performance across the different testing scenarios. Evaluation against SYNOP reports and Himawari satellite imagery showed that the predictions were more consistent under atmospheric conditions with distinct characteristics, whereas discrepancies occurred more frequently under transitional atmospheric conditions. Comparison during the 2024 testing period showed that SVM had a rainfall detection capability relatively comparable to the BMKG NWP-based forecast, although NWP provided a more balanced ability to distinguish between clear and rainy conditions. These findings indicate that radiosonde-based machine learning using atmospheric stability indices can serve as a supporting approach for weather condition prediction in tropical Cilacap.

Kata kunci

Pembimbing 1

Pembimbing 2

Pembimbing 3

Tahun

Jumlah Halaman