Pemodelan Ketebalan Gambut Menggunakan DTM LiDAR dan Data Multisumber Dengan Pendekatan Machine Learning di KHG Sungai Mendahara-Sungai Batanghari

Fahri Reza, Dr. Ir. Bambang Kun Cahyono, S.T., M.Sc., IPU.

2025 | Tesis | S2 Teknik Geomatika

Ketebalan gambut merupakan parameter krusial dalam pengelolaan ekosistem gambut yang memerlukan informasi akurat. Metode konvensional seperti pengeboran masih digunakan karena akurasinya, namun memerlukan biaya tinggi dan waktu yang lama untuk area luas. Pemodelan ketebalan gambut menggunakan integrasi data DTM LiDAR, citra satelit Planetscope, data regional, dan data global menawarkan solusi efisien dan akurat. Penelitian ini bertujuan memodelkan prediksi ketebalan gambut menggunakan DTM dari LiDAR dan data multisumber dengan pendekatan machine learning di Kesatuan Hidrologis Gambut (KHG) Sungai Mendahara-Sungai Batanghari.

Penelitian dilakukan di area seluas 201.383 hektar menggunakan 358 data pengeboran ketebalan gambut yang telah dibersihkan dari missing values dan outlier dari total 1.183 titik awal. Sebanyak 17 kovariat digunakan meliputi parameter hidrotopografi DTM dari LiDAR (elevation, slope, aspect, curvature, TPI, TWI, SPI, MRVBF, geomorphons), indeks vegetasi NDVI, data tutupan lahan, data geologi, jarak ke sungai terdekat, ketebalan tanah global, ketebalan gambut global, karbon organik tanah global, dan model gangguan gravitasi. Seleksi fitur menggunakan algoritma Boruta, pembagian data dengan rasio 80:20 menggunakan stratified sampling, dan validasi model dengan 5-fold cross validation. Tiga algoritma machine learning (Random Forest, Cubist, dan XGBoost) dibandingkan performanya melalui hyperparameter tuning berdasarkan metrik evaluasi RMSE, MAE, dan R².

Hasil seleksi fitur mengidentifikasi 11 kovariat signifikan, meliputi 10 kovariat terkonfirmasi (Elevation, MRVBF, LandCover, Global_PeatThickness, NDVI, Gravity_Disturbance, Geologi, Global_Soil, NearestToRiver, TPI) dan 1 kovariat tentative (Geomorphons). Evaluasi menunjukkan Cubist memberikan nilai terbaik dengan RMSE 70,88 cm, MAE 50,20 cm, dan R² 0,82, lebih unggul dibandingkan XGBoost (RMSE 82,25 cm, R² 0,75) dan Random Forest (RMSE 86,71 cm, R² 0,73). Model Cubist menghasilkan peta spasial dengan pola distribusi ketebalan tertinggi terkonsentrasi di bagian tengah dan timur wilayah penelitian. Analisis variable importance mengidentifikasi Elevation sebagai kovariat paling berpengaruh dengan kontribusi 23,80?lam prediksi ketebalan gambut.

Peat thickness is a crucial parameter in peatland ecosystem management that requires accurate information. Conventional methods such as coring are still used due to their accuracy, but they require high costs and considerable time for large areas. Peat thickness modeling using the integration of LiDAR DTM data, Planetscope satellite imagery, regional data, and global data offers an efficient and accurate solution. This study aims to model peat thickness estimation using LiDAR DTM and multisource data with a machine learning approach in the Peat Hydrological Unit (PHU) of Sungai Mendahara-Sungai Batanghari.

The study was conducted in an area of 201,383 hectares using 358 peat thickness coring data that had been cleaned from missing values and outliers from a total of 1,183 initial points. A total of 17 covariates were used, comprising LiDAR DTM hydrotopographic parameters (elevation, slope, aspect, curvature, TPI, TWI, SPI, MRVBF, geomorphons), NDVI vegetation index, land cover data, geological data, distance to nearest river, global soil thickness, global peat thickness, global soil organic carbon, and gravity disturbance model. Feature selection used the Boruta algorithm, data splitting with an 80:20 ratio using stratified sampling, and model validation with 5-fold cross-validation. Three machine learning algorithms (Random Forest, Cubist, and XGBoost) were compared for their performance through hyperparameter tuning based on evaluation metrics of RMSE, MAE, and R².

Feature selection results identified 11 significant covariates, comprising 10 confirmed covariates (Elevation, MRVBF, LandCover, Global_PeatThickness, NDVI, Gravity_Disturbance, Geology, Global_Soil, NearestToRiver, TPI) and 1 tentative covariate (Geomorphons). Evaluation showed that Cubist provided the best performance with an RMSE of 70.88 cm, MAE of 50.20 cm, and R² of 0.82, outperforming XGBoost (RMSE 82.25 cm, R² 0.75) and Random Forest (RMSE 86.71 cm, R² 0.73). The Cubist model produced a spatial map with the highest thickness distribution pattern concentrated in the central and eastern parts of the study area. Variable importance analysis identified Elevation as the most influential covariate with a contribution of 23.80% in peat thickness prediction.

Kata Kunci : ketebalan gambut, DTM LiDAR, machine learning, KHG Sungai Mendahara-Sungai Batanghari, pemodelan gambut.