Melola Discovery

← Melola Discovery

Open access Journal SINTA 2

Adaptive Ensemble CNN–BiLSTM untuk Integrasi Citra Awan dan Data Meteorologi BMKG dalam Prediksi Cuaca Jangka Pendek di Manokwari

Jurnal Teknologi Informasi dan Ilmu Komputer (JTIIK) · Vol. 13 No. 3 (2026) · pp. 757–772

DOI
10.25126/jtiik.2026133
ISSN
2528-6579, 2355-7699
Published
2026-07-01
Publisher
Fakultas Ilmu Komputer, Universitas Brawijaya
Language
id
License
creativecommons.org/licenses/by-sa/4.0

Abstract

Indonesian

Penelitian ini mengusulkan model Adaptive Ensemble CNN–BiLSTM untuk meningkatkan konsistensi dan performa rata-rata prakiraan cuaca jangka pendek (1–3 jam) di wilayah tropis Manokwari, Papua Barat. Model Adaptive Ensemble dibangun dengan mengintegrasikan citra awan multisumber, data meteorologi BMKG, serta metadata waktu dan lokasi. Convolutional Neural Network (CNN) digunakan untuk mengekstraksi fitur spasial, sedangkan Bidirectional Long Short-Term Memory (BiLSTM) memodelkan dinamika temporal deret waktu atmosfer. Mekanisme Adaptive Ensemble menggabungkan prediksi CNN dan BiLSTM dengan bobot dinamis, sehingga performa rata-rata meningkat signifikan dibandingkan baseline sederhana maupun Conditional Ensemble. Evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan macro F1 pada horizon H+1, H+2, dan H+3, serta uji statistik (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, dan Balanced Accuracy). Hasil menunjukkan bahwa BiLSTM unggul dalam global accuracy (88–89%, macro F1 0.88), sementara Adaptive Ensemble menghasilkan accuracy 87–90% dengan macro F1 0.84–0.88. Uji paired t-test signifikan (p = 0.017) menegaskan bahwa Adaptive Ensemble unggul secara rata-rata, dengan distribusi kesalahan yang lebih konsisten lintas horizon. Analisis metadata memperlihatkan bahwa Adaptive Ensemble lebih unggul pada siang hari (accuracy 0.90–0.94, F1 0.83–0.92), sedangkan Conditional Ensemble lebih stabil pada malam hari (accuracy 0.81–0.82, F1 0.80–0.81). Evaluasi antar kelurahan menunjukkan mayoritas accuracy Adaptive Ensemble berada pada kisaran 0.80–0.92, dengan pengecualian Udopi yang lebih rendah (0.652, F1 0.558). Hal ini menandakan distribusi relatif seimbang tanpa bias ekstrem, meskipun jumlah data memengaruhi variasi performa. Kontribusi utama penelitian ini adalah integrasi multimodal CNN–BiLSTM dengan Adaptive Ensemble serta evaluasi berbasis metadata, yang menghasilkan prakiraan cuaca lokal lebih akurat secara rata-rata sekaligus relevan lintas waktu dan lokasi. Abstract This study proposes an Adaptive Ensemble CNN–BiLSTM model to improve the consistency and average performance of short-term weather forecasts (1–3 hours) in the tropical region of Manokwari, West Papua. The Adaptive Ensemble model is built by integrating multi-source cloud imagery, BMKG meteorological data, and time and location metadata. Convolutional Neural Network (CNN) is used to extract spatial features, while Bidirectional Long Short-Term Memory (BiLSTM) models the temporal dynamics of atmospheric time series. The Adaptive Ensemble mechanism combines CNN and BiLSTM predictions with dynamic weights, resulting in a significant improvement in average performance compared to the simple baseline and the Conditional Ensemble. Evaluation is carried out using accuracy, precision, recall, and macro F1 metrics at horizons H+1, H+2, and H+3, as well as statistical tests (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, and Balanced Accuracy). The results show that BiLSTM excels in global accuracy (88–89%, macro F1 0.88), while Adaptive Ensemble produces 87–90% accuracy with macro F1 0.84–0.88. A significant paired t-test (p = 0.017) confirms that Adaptive Ensemble is superior on average, with a more consistent error distribution across horizons. Metadata analysis shows that Adaptive Ensemble is superior during the day (accuracy 0.90–0.94, F1 0.83–0.92), while Conditional Ensemble is more stable at night (accuracy 0.81–0.82, F1 0.80–0.81). Evaluation between villages shows that the majority of Adaptive Ensemble accuracy is in the range of 0.80–0.92, with the exception of Udopi which is lower (0.652, F1 0.558). This indicates a relatively balanced distribution without extreme bias, although the amount of data influences performance variation. The main contribution of this research is the integration of multimodal CNN–BiLSTM with Adaptive Ensemble and metadata-based evaluation, which results in local weather forecasts that are more accurate on average and relevant across time and locations.

English

Penelitian ini mengusulkan model Adaptive Ensemble CNN–BiLSTM untuk meningkatkan konsistensi dan performa rata-rata prakiraan cuaca jangka pendek (1–3 jam) di wilayah tropis Manokwari, Papua Barat. Model Adaptive Ensemble dibangun dengan mengintegrasikan citra awan multisumber, data meteorologi BMKG, serta metadata waktu dan lokasi. Convolutional Neural Network (CNN) digunakan untuk mengekstraksi fitur spasial, sedangkan Bidirectional Long Short-Term Memory (BiLSTM) memodelkan dinamika temporal deret waktu atmosfer. Mekanisme Adaptive Ensemble menggabungkan prediksi CNN dan BiLSTM dengan bobot dinamis, sehingga performa rata-rata meningkat signifikan dibandingkan baseline sederhana maupun Conditional Ensemble. Evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan macro F1 pada horizon H+1, H+2, dan H+3, serta uji statistik (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, dan Balanced Accuracy). Hasil menunjukkan bahwa BiLSTM unggul dalam global accuracy (88–89%, macro F1 0.88), sementara Adaptive Ensemble menghasilkan accuracy 87–90% dengan macro F1 0.84–0.88. Uji paired t-test signifikan (p = 0.017) menegaskan bahwa Adaptive Ensemble unggul secara rata-rata, dengan distribusi kesalahan yang lebih konsisten lintas horizon. Analisis metadata memperlihatkan bahwa Adaptive Ensemble lebih unggul pada siang hari (accuracy 0.90–0.94, F1 0.83–0.92), sedangkan Conditional Ensemble lebih stabil pada malam hari (accuracy 0.81–0.82, F1 0.80–0.81). Evaluasi antar kelurahan menunjukkan mayoritas accuracy Adaptive Ensemble berada pada kisaran 0.80–0.92, dengan pengecualian Udopi yang lebih rendah (0.652, F1 0.558). Hal ini menandakan distribusi relatif seimbang tanpa bias ekstrem, meskipun jumlah data memengaruhi variasi performa. Kontribusi utama penelitian ini adalah integrasi multimodal CNN–BiLSTM dengan Adaptive Ensemble serta evaluasi berbasis metadata, yang menghasilkan prakiraan cuaca lokal lebih akurat secara rata-rata sekaligus relevan lintas waktu dan lokasi. Abstract This study proposes an Adaptive Ensemble CNN–BiLSTM model to improve the consistency and average performance of short-term weather forecasts (1–3 hours) in the tropical region of Manokwari, West Papua. The Adaptive Ensemble model is built by integrating multi-source cloud imagery, BMKG meteorological data, and time and location metadata. Convolutional Neural Network (CNN) is used to extract spatial features, while Bidirectional Long Short-Term Memory (BiLSTM) models the temporal dynamics of atmospheric time series. The Adaptive Ensemble mechanism combines CNN and BiLSTM predictions with dynamic weights, resulting in a significant improvement in average performance compared to the simple baseline and the Conditional Ensemble. Evaluation is carried out using accuracy, precision, recall, and macro F1 metrics at horizons H+1, H+2, and H+3, as well as statistical tests (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, and Balanced Accuracy). The results show that BiLSTM excels in global accuracy (88–89%, macro F1 0.88), while Adaptive Ensemble produces 87–90% accuracy with macro F1 0.84–0.88. A significant paired t-test (p = 0.017) confirms that Adaptive Ensemble is superior on average, with a more consistent error distribution across horizons. Metadata analysis shows that Adaptive Ensemble is superior during the day (accuracy 0.90–0.94, F1 0.83–0.92), while Conditional Ensemble is more stable at night (accuracy 0.81–0.82, F1 0.80–0.81). Evaluation between villages shows that the majority of Adaptive Ensemble accuracy is in the range of 0.80–0.92, with the exception of Udopi which is lower (0.652, F1 0.558). This indicates a relatively balanced distribution without extreme bias, although the amount of data influences performance variation. The main contribution of this research is the integration of multimodal CNN–BiLSTM with Adaptive Ensemble and metadata-based evaluation, which results in local weather forecasts that are more accurate on average and relevant across time and locations.

Keywords

Collaboration network of the authors

3 authors · 3 co-authorship links

Hover or tap a name to highlight its co-authors. Double-click to open the author's page.

  • Affiliation unknown
  • Different organisations
  • Same organisation
  • Affiliation unknown
  • Thicker line = more joint articles · bigger circle = more articles
Show as table
Co-authorOrganisationJoint articles

Affiliations come from OpenAlex; journal OAI-PMH metadata carries none. Authors without an OpenAlex ID are matched by name, which can merge or split people.

Metadata source: OAI-PMH (metadata jurnal) https://jtiik.ub.ac.id/index.php/jtiik/en/oai, first harvested 27 Oct 2026, last seen 15 Nov 2026.
Original metadata (oai_dc)
<oai_dc:dc xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:oai_dc="http://www.openarchives.org/OAI/2.0/oai_dc/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/oai_dc/  http://www.openarchives.org/OAI/2.0/oai_dc.xsd">
	<dc:title xml:lang="id">Adaptive Ensemble CNN–BiLSTM untuk Integrasi Citra Awan dan Data Meteorologi BMKG dalam Prediksi Cuaca Jangka Pendek di Manokwari</dc:title>
	<dc:title xml:lang="en">Adaptive Ensemble CNN–BiLSTM untuk Integrasi Citra Awan dan Data Meteorologi BMKG dalam Prediksi Cuaca Jangka Pendek di Manokwari</dc:title>
	<dc:creator xml:lang="en">Nur Fauzi, Ridho</dc:creator>
	<dc:creator xml:lang="en">P. P Naibaho, Julius</dc:creator>
	<dc:creator xml:lang="en"> De Kweldju, Alex</dc:creator>
	<dc:creator xml:lang="id">Nur Fauzi, Ridho</dc:creator>
	<dc:creator xml:lang="id">P. P Naibaho, Julius</dc:creator>
	<dc:creator xml:lang="id"> De Kweldju, Alex</dc:creator>
	<dc:subject xml:lang="en">CNN</dc:subject>
	<dc:subject xml:lang="en">BiLSTM</dc:subject>
	<dc:subject xml:lang="en">Conditional Ensamble</dc:subject>
	<dc:subject xml:lang="en">Cloud Imagery</dc:subject>
	<dc:subject xml:lang="en">BMKG</dc:subject>
	<dc:subject xml:lang="en">Weather Forecasting</dc:subject>
	<dc:subject xml:lang="id">BiLSTM</dc:subject>
	<dc:subject xml:lang="id">citra awan</dc:subject>
	<dc:subject xml:lang="id">BMKG</dc:subject>
	<dc:subject xml:lang="id">CNN</dc:subject>
	<dc:subject xml:lang="id">Conditional Ensamble</dc:subject>
	<dc:subject xml:lang="id">Prakiraan Cuaca</dc:subject>
	<dc:description xml:lang="id">Penelitian ini mengusulkan model Adaptive Ensemble CNN–BiLSTM untuk meningkatkan konsistensi dan performa rata-rata prakiraan cuaca jangka pendek (1–3 jam) di wilayah tropis Manokwari, Papua Barat. Model Adaptive Ensemble dibangun dengan mengintegrasikan citra awan multisumber, data meteorologi BMKG, serta metadata waktu dan lokasi. Convolutional Neural Network (CNN) digunakan untuk mengekstraksi fitur spasial, sedangkan Bidirectional Long Short-Term Memory (BiLSTM) memodelkan dinamika temporal deret waktu atmosfer. Mekanisme Adaptive Ensemble menggabungkan prediksi CNN dan BiLSTM dengan bobot dinamis, sehingga performa rata-rata meningkat signifikan dibandingkan baseline sederhana maupun Conditional Ensemble. Evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan macro F1 pada horizon H+1, H+2, dan H+3, serta uji statistik (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, dan Balanced Accuracy). Hasil menunjukkan bahwa BiLSTM unggul dalam global accuracy (88–89%, macro F1 0.88), sementara Adaptive Ensemble menghasilkan accuracy 87–90% dengan macro F1 0.84–0.88. Uji paired t-test signifikan (p = 0.017) menegaskan bahwa Adaptive Ensemble unggul secara rata-rata, dengan distribusi kesalahan yang lebih konsisten lintas horizon. Analisis metadata memperlihatkan bahwa Adaptive Ensemble lebih unggul pada siang hari (accuracy 0.90–0.94, F1 0.83–0.92), sedangkan Conditional Ensemble lebih stabil pada malam hari (accuracy 0.81–0.82, F1 0.80–0.81). Evaluasi antar kelurahan menunjukkan mayoritas accuracy Adaptive Ensemble berada pada kisaran 0.80–0.92, dengan pengecualian Udopi yang lebih rendah (0.652, F1 0.558). Hal ini menandakan distribusi relatif seimbang tanpa bias ekstrem, meskipun jumlah data memengaruhi variasi performa. Kontribusi utama penelitian ini adalah integrasi multimodal CNN–BiLSTM dengan Adaptive Ensemble serta evaluasi berbasis metadata, yang menghasilkan prakiraan cuaca lokal lebih akurat secara rata-rata sekaligus relevan lintas waktu dan lokasi.
 
Abstract
This study proposes an Adaptive Ensemble CNN–BiLSTM model to improve the consistency and average performance of short-term weather forecasts (1–3 hours) in the tropical region of Manokwari, West Papua. The Adaptive Ensemble model is built by integrating multi-source cloud imagery, BMKG meteorological data, and time and location metadata. Convolutional Neural Network (CNN) is used to extract spatial features, while Bidirectional Long Short-Term Memory (BiLSTM) models the temporal dynamics of atmospheric time series. The Adaptive Ensemble mechanism combines CNN and BiLSTM predictions with dynamic weights, resulting in a significant improvement in average performance compared to the simple baseline and the Conditional Ensemble. Evaluation is carried out using accuracy, precision, recall, and macro F1 metrics at horizons H+1, H+2, and H+3, as well as statistical tests (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, and Balanced Accuracy). The results show that BiLSTM excels in global accuracy (88–89%, macro F1 0.88), while Adaptive Ensemble produces 87–90% accuracy with macro F1 0.84–0.88. A significant paired t-test (p = 0.017) confirms that Adaptive Ensemble is superior on average, with a more consistent error distribution across horizons. Metadata analysis shows that Adaptive Ensemble is superior during the day (accuracy 0.90–0.94, F1 0.83–0.92), while Conditional Ensemble is more stable at night (accuracy 0.81–0.82, F1 0.80–0.81). Evaluation between villages shows that the majority of Adaptive Ensemble accuracy is in the range of 0.80–0.92, with the exception of Udopi which is lower (0.652, F1 0.558). This indicates a relatively balanced distribution without extreme bias, although the amount of data influences performance variation. The main contribution of this research is the integration of multimodal CNN–BiLSTM with Adaptive Ensemble and metadata-based evaluation, which results in local weather forecasts that are more accurate on average and relevant across time and locations.</dc:description>
	<dc:description xml:lang="en">Penelitian ini mengusulkan model Adaptive Ensemble CNN–BiLSTM untuk meningkatkan konsistensi dan performa rata-rata prakiraan cuaca jangka pendek (1–3 jam) di wilayah tropis Manokwari, Papua Barat. Model Adaptive Ensemble dibangun dengan mengintegrasikan citra awan multisumber, data meteorologi BMKG, serta metadata waktu dan lokasi. Convolutional Neural Network (CNN) digunakan untuk mengekstraksi fitur spasial, sedangkan Bidirectional Long Short-Term Memory (BiLSTM) memodelkan dinamika temporal deret waktu atmosfer. Mekanisme Adaptive Ensemble menggabungkan prediksi CNN dan BiLSTM dengan bobot dinamis, sehingga performa rata-rata meningkat signifikan dibandingkan baseline sederhana maupun Conditional Ensemble. Evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan macro F1 pada horizon H+1, H+2, dan H+3, serta uji statistik (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, dan Balanced Accuracy). Hasil menunjukkan bahwa BiLSTM unggul dalam global accuracy (88–89%, macro F1 0.88), sementara Adaptive Ensemble menghasilkan accuracy 87–90% dengan macro F1 0.84–0.88. Uji paired t-test signifikan (p = 0.017) menegaskan bahwa Adaptive Ensemble unggul secara rata-rata, dengan distribusi kesalahan yang lebih konsisten lintas horizon. Analisis metadata memperlihatkan bahwa Adaptive Ensemble lebih unggul pada siang hari (accuracy 0.90–0.94, F1 0.83–0.92), sedangkan Conditional Ensemble lebih stabil pada malam hari (accuracy 0.81–0.82, F1 0.80–0.81). Evaluasi antar kelurahan menunjukkan mayoritas accuracy Adaptive Ensemble berada pada kisaran 0.80–0.92, dengan pengecualian Udopi yang lebih rendah (0.652, F1 0.558). Hal ini menandakan distribusi relatif seimbang tanpa bias ekstrem, meskipun jumlah data memengaruhi variasi performa. Kontribusi utama penelitian ini adalah integrasi multimodal CNN–BiLSTM dengan Adaptive Ensemble serta evaluasi berbasis metadata, yang menghasilkan prakiraan cuaca lokal lebih akurat secara rata-rata sekaligus relevan lintas waktu dan lokasi.
 
Abstract
This study proposes an Adaptive Ensemble CNN–BiLSTM model to improve the consistency and average performance of short-term weather forecasts (1–3 hours) in the tropical region of Manokwari, West Papua. The Adaptive Ensemble model is built by integrating multi-source cloud imagery, BMKG meteorological data, and time and location metadata. Convolutional Neural Network (CNN) is used to extract spatial features, while Bidirectional Long Short-Term Memory (BiLSTM) models the temporal dynamics of atmospheric time series. The Adaptive Ensemble mechanism combines CNN and BiLSTM predictions with dynamic weights, resulting in a significant improvement in average performance compared to the simple baseline and the Conditional Ensemble. Evaluation is carried out using accuracy, precision, recall, and macro F1 metrics at horizons H+1, H+2, and H+3, as well as statistical tests (McNemar, paired t-test, Wilcoxon, Cohen’s Kappa, and Balanced Accuracy). The results show that BiLSTM excels in global accuracy (88–89%, macro F1 0.88), while Adaptive Ensemble produces 87–90% accuracy with macro F1 0.84–0.88. A significant paired t-test (p = 0.017) confirms that Adaptive Ensemble is superior on average, with a more consistent error distribution across horizons. Metadata analysis shows that Adaptive Ensemble is superior during the day (accuracy 0.90–0.94, F1 0.83–0.92), while Conditional Ensemble is more stable at night (accuracy 0.81–0.82, F1 0.80–0.81). Evaluation between villages shows that the majority of Adaptive Ensemble accuracy is in the range of 0.80–0.92, with the exception of Udopi which is lower (0.652, F1 0.558). This indicates a relatively balanced distribution without extreme bias, although the amount of data influences performance variation. The main contribution of this research is the integration of multimodal CNN–BiLSTM with Adaptive Ensemble and metadata-based evaluation, which results in local weather forecasts that are more accurate on average and relevant across time and locations.</dc:description>
	<dc:publisher xml:lang="en">Fakultas Ilmu Komputer, Universitas Brawijaya</dc:publisher>
	<dc:date>2026-07-01</dc:date>
	<dc:type>info:eu-repo/semantics/article</dc:type>
	<dc:type>info:eu-repo/semantics/publishedVersion</dc:type>
	<dc:format>application/pdf</dc:format>
	<dc:identifier>https://jtiik.ub.ac.id/index.php/jtiik/article/view/10866</dc:identifier>
	<dc:identifier>10.25126/jtiik.2026133</dc:identifier>
	<dc:source xml:lang="id">Jurnal Teknologi Informasi dan Ilmu Komputer; Vol 13 No 3: Juni 2026; 757-772</dc:source>
	<dc:source xml:lang="en">Jurnal Teknologi Informasi dan Ilmu Komputer; Vol. 13 No. 3: Juni 2026; 757-772</dc:source>
	<dc:source>2528-6579</dc:source>
	<dc:source>2355-7699</dc:source>
	<dc:source>10.25126/jtiik.2026133</dc:source>
	<dc:language>id</dc:language>
	<dc:relation>https://jtiik.ub.ac.id/index.php/jtiik/article/view/10866/1608</dc:relation>
	<dc:rights xml:lang="en">Copyright (c) 2026 Ridho Nur Fauzi, Julius P. P Naibaho, Alex  De Kweldju</dc:rights>
	<dc:rights xml:lang="en">https://creativecommons.org/licenses/by-sa/4.0</dc:rights>
</oai_dc:dc>