Laporkan Masalah

Machine Learning-Based Acoustic Analysis Model Combining Mel-Frequency Cepstral Coefficients (MFCC) With Acoustic Indices For Forest Sound Classification At Wanagama Forest, Indonesia

ALIYU SALEH MAGAJI, Ir. Sentagi Sesotya Utami, S.T., M.Sc., Ph.D., IPU

2026 | Tesis | MAGISTER TEKNIK FISIKA

Klasifikasi lanskap suara hutan menjadi Biophony, Anthrophony, dan Geophony menjadi landasan pemantauan akustik ekosistem hutan. Namun, deskriptor spektral dan indeks akustik ekologi sebagian besar telah digunakan secara terpisah, dan integrasinya jarang diteliti melalui perbandingan terkontrol dalam konteks hutan tropis. Penelitian ini mengembangkan dan mengevaluasi model analisis akustik berbasis pembelajaran mesin yang menggabungkan Mel-Frequency Cepstral Coefficients (MFCC) dengan indeks akustik untuk mengklasifikasikan ketiga komponen lanskap suara tersebut di Hutan Wanagama, Indonesia, serta untuk menetapkan nilai komparatif dari representasi spektral, berbasis indeks, dan terintegrasi. Sebuah dataset seimbang yang terdiri dari 3.000 rekaman suara hutan dikarakterisasi menggunakan fitur MFCC pada tiga dimensi dan enam indeks akustik, sehingga membentuk tujuh set fitur. Set-set fitur ini dievaluasi menggunakan klasifikator Regresi Logistik, Hutan Acak, dan Mesin Vektor Pendukung (SVM) melalui validasi silang lima lipat yang terstratifikasi, dengan skor F1 sebagai metrik utama. Perbedaan antar konfigurasi diuji menggunakan uji Friedman, diikuti dengan perbandingan Wilcoxon signed-rank yang dikoreksi dengan metode Bonferroni. Hasil penelitian menunjukkan bahwa representasi berbasis MFCC dan representasi terintegrasi mengandung informasi diskriminatif yang diperlukan untuk klasifikasi yang andal, sedangkan indeks akustik yang digunakan secara terpisah secara konsisten merupakan kelompok fitur terlemah. Representasi terintegrasi MFCC-26 dan indeks akustik yang dipadukan dengan Random Forest mencapai skor F1 tertinggi dan paling stabil.

Forest soundscape classification into Biophony, Anthrophony, and Geophony underpins the acoustic monitoring of forest ecosystems. Spectral descriptors and ecological acoustic indices, however, have largely been used in isolation, and their integration has rarely been examined under controlled comparison in a tropical-forest context. This study developed and evaluated a machine learning-based acoustic analysis model that combines Mel-Frequency Cepstral Coefficients (MFCC) with acoustic indices to classify these three soundscape components at Wanagama Forest, Indonesia, and to establish the comparative value of spectral, index-based, and integrated representations. A balanced dataset of 3,000 forest sound recordings was characterised using MFCC features at three dimensionalities and six acoustic indices, forming seven feature sets. These were evaluated across Logistic Regression, Random Forest, and Support Vector Machine classifiers using stratified fivefold cross-validation, with the F1-score as the primary metric. Differences among configurations were tested using the Friedman test, followed by Bonferroni-corrected Wilcoxon signed-rank comparisons. The results show that MFCC-based and integrated representations carry the discriminative information needed for reliable classification, whereas acoustic indices used alone are consistently the weakest of the feature groups. The integrated MFCC-26 and acoustic-index representation with Random Forest achieved the highest and most stable performance (mean F1 = 0.9923); ensemble and margin-based classifiers were robust across feature sets, while the linear classifier remained the most dependent on the chosen representation. The Friedman test rejected the null hypothesis of equal performance among configurations, but the corrected post-hoc comparisons showed no significant pairwise difference. Integration therefore did not significantly improve accuracy over MFCC features alone; its main benefit was greater stability. A supplementary plot-level comparison further showed that the NDSI and BI differed significantly between the two deployment plots, consistent with their designed anthropogenic disturbance gradient, whereas the biologically driven ACI did not. The study contributes a controlled comparative evaluation and a reproducible framework that clarify the respective roles of MFCC-only, acoustic-index-only, and integrated representations for forest sound classification within the Wanagama context.

Kata Kunci : Forest sound classification, Mel-Frequency Cepstral Coefficients (MFCC), Acoustic indices, Machine learning, Soundscape analysis

  1. S2-2026-550374-abstract.pdf  
  2. S2-2026-550374-bibliography.pdf  
  3. S2-2026-550374-tableofcontent.pdf  
  4. S2-2026-550374-title.pdf