Analisis Pola Upwelling dan Penerapan Model LSTM untuk Deteksi di WPP-NRI 572 Menggunakan Data SST dan Chl-a

Abstract

Fenomena upwelling di Wilayah Pengelolaan Perikanan Negara Republik Indonesia (WPP-NRI) 572 berperan penting dalam mendukung produktivitas laut melalui suplai massa air kaya nutrien dari lapisan dalam ke permukaan. Ciri khasnya adalah penurunan suhu permukaan laut (SST) dan peningkatan klorofil-a (Chl-a) yang mendorong pertumbuhan fitoplankton. Penelitian ini bertujuan mengidentifikasi karakteristik spasial-temporal upwelling, mengevaluasi metode Interquartile Range (IQR) dan time-window untuk mendeteksi intensitas upwelling, serta mengembangkan model Long Short-Term Memory (LSTM) untuk klasifikasi dan prediksi berdasarkan data SST dan Chl-a periode 2007–2017. Hasil menunjukkan upwelling terjadi musiman pada angin monsun tenggara (Juni–September), dengan intensitas tinggi di pesisir barat Sumatra selatan dan Selat Sunda. Metode IQR dan time-window efektif menyaring anomali dan mendeteksi kejadian signifikan. Model LSTM terbaik (5 hidden layer, 512 neuron) mencapai akurasi 92,14% dan F1-score 81,49%, dengan 77 lokasi akurasi =90%. Prediksi regresi SST terbaik (1 hidden layer, 256 neuron) dengan R² 0,88, RMSE 0,28 °C; prediksi Chl-a terbaik (1 hidden layer, 512 neuron) dengan R² 0,77, RMSE 0,04 mg/m³. Integrasi output model regresi dalam klasifikasi menunjukkan performa yang bervariasi di beberapa periode waktu yang belum optimal dalam menangkap pola spasial dan temporal fenomena upwelling secara akurat. Namun, secara keseluruhan pendekatan LSTM ini berpotensi besar untuk sistem pemantauan laut dan pengelolaan perikanan berkelanjutan di WPP-NRI 572.

The upwelling phenomenon in the Fisheries Management Area of the Republic of Indonesia (WPP-NRI) 572 plays a crucial role in supporting marine productivity by supplying nutrient-rich deep water masses to the surface. It is characterized by a decrease in sea surface temperature (SST) and an increase in chlorophyll-a (Chl-a) concentration, which promotes phytoplankton growth. This study aims to identify the spatial–temporal characteristics of upwelling, evaluate the Interquartile Range (IQR) and time-window methods for detecting upwelling intensity, and develop a Long Short-Term Memory (LSTM) model for classification and prediction based on SST and Chl-a data from 2007 to 2017. The results indicate that upwelling occurs seasonally during the southeast monsoon (June–September), with high intensity observed along the southern west coast of Sumatra and the Sunda Strait. The IQR and time-window methods effectively filtered anomalies and detected significant events. The best-performing LSTM model (5 hidden layers, 512 neurons) achieved an accuracy of 92.14% and an F1-score of 81.49%, with 77 locations recording accuracy =90%. The best SST regression performance (1 hidden layer, 256 neurons) yielded an R² of 0.88 and an RMSE of 0.28 °C, while the best Chl-a regression performance (1 hidden layer, 512 neurons) achieved an R² of 0.77 and an RMSE of 0.04 mg/m³. Integrating the regression model outputs into classification showed varying performance across different time periods, with some limitations in accurately capturing the spatial and temporal patterns of upwelling. Nevertheless, the overall LSTM-based approach demonstrates significant potential for marine monitoring systems and sustainable fisheries management in WPP-NRI 572.