Electronic Theses and Dissertation

Kejadian upwelling di danau tropis menimbulkan risiko ekologis dan ekonomi yang signifikan terhadap sistem budidaya perikanan, terutama melalui kematian massal ikan. Penelitian sebelumnya umumnya berfokus pada deteksi kejadian dan jarang mengintegrasikan representasi ketidakpastian, dinamika temporal, dan klasifikasi nonlinier dalam satu kerangka prediksi dini. Penelitian ini bertujuan mengembangkan kerangka prediktif terintegrasi untuk mengidentifikasi potensi kejadian upwelling di Danau Maninjau, Indonesia. Penelitian ini menggunakan desain observasional berbasis pemodelan prediktif dengan data meteorologi harian periode 2001–2025 yang diperoleh dari NASA POWER pada koordinat 0,32°LS dan 100,19°BT. Variabel yang dianalisis meliputi radiasi matahari, tekanan permukaan, temperatur, kecepatan angin, dan indeks kejernihan langit. Klasterisasi Fuzzy C-Shells digunakan untuk mengidentifikasi stabilitas kondisi iklim yang merepresentasikan potensi upwelling. Model klasifikasi CatBoost dengan optimasi hyperparameter berbasis grid search diterapkan untuk prediksi kejadian, sedangkan model Vector Autoregressive (VAR) dengan orde lag optimal 31 (berdasarkan AIC dan FPE) digunakan untuk peramalan multivariat. Evaluasi kinerja dilakukan menggunakan akurasi, balanced accuracy, precision, recall, F1-score, ROC-AUC, serta uji stabilitas VAR. Klasterisasi fuzzy mengidentifikasi dua rezim iklim utama yang merepresentasikan kondisi berpotensi dan tidak berpotensi upwelling. Model CatBoost teroptimasi mencapai akurasi 0,91, precision 0,89, recall 0,87, F1-score 0,88, dan AUC 0,93 pada data uji. Model VAR(31) memenuhi kondisi stabilitas dengan seluruh akar karakteristik berada di dalam unit circle, menunjukkan dinamika temporal yang kuat. Kecepatan angin dan radiasi matahari teridentifikasi sebagai faktor dominan pembentukan kondisi upwelling. Kerangka terintegrasi FCS, CatBoost, dan VAR meningkatkan performa prediksi dan interpretabilitas dibandingkan pendekatan tunggal.

Upwelling events in tropical lakes pose significant ecological and economic risks to aquaculture systems, particularly through mass fish mortality. Existing studies primarily focus on detection rather than early prediction and rarely integrate uncertainty representation, temporal dynamics, and nonlinear classification. This study addresses this gap by developing an integrated predictive framework for upwelling potential in Lake Maninjau, Indonesia. An observational modeling study was conducted using daily meteorological data from 2001 to 2025 obtained from NASA POWER at coordinates 0.32°S and 100.19°E. Variables included solar radiation, surface pressure, temperature, wind speed, and sky clearness index. Fuzzy C-Shells clustering was applied to identify climate condition clusters representing upwelling potential. A CatBoost classifier with grid-search hyperparameter tuning was used for classification, while a multivariate Vector Autoregressive (VAR) model with optimal lag 31 (based on AIC and FPE) was used for temporal forecasting. Model performance was evaluated using accuracy, balanced accuracy, precision, recall, F1-score, ROC-AUC, and VAR stability diagnostics. Fuzzy clustering identified two dominant climate stability corresponding to potential and non-potential upwelling conditions. The optimized CatBoost model achieved accuracy of 0.91, precision of 0.89, recall of 0.87, F1-score of 0.88, and AUC of 0.93 on the test dataset. The VAR(31) model satisfied stability conditions with all characteristic roots inside the unit circle, indicating robust temporal dynamics. Wind speed and solar radiation were identified as dominant drivers of upwelling conditions. The integrated FCS, CatBoost, and VAR framework improves predictive performance and interpretability compared to single-method approaches.