Pembangunan Algoritma Decision Tree of Machine Learning dalam Deteksi Deforestasi Mangrove di Kabupaten Langkat

Abstract

Penelitian ini menerangkan tentang bagaimana mengembangkan metode pendeteksi perubahan tutupan mangrove berbasis pendekatan pembelajar mesin (machine learning). Penelitian dilakukan terhadap kejadian deforestasi yang terjadi di Kabupaten Langkat pada periode mulai tahun 2018 sampai dengan tahun 2024. Citra perubahan dibuat menggunakan metode image differencing. Model perubahan tutupan mangrove dibangun menggunakan kombinasi peubah citra hasil image differencing dan peubah sosio-geo-biofisik. Penelitian ini menunjukkan bahwa model terbaik dibangun menggunakan kombinasi peubah image differencing dan sosio-geo-biofisik dengan penghalusan citra pasca klasifikasi menghasilkan overall accuracy sebesar 95.1%, kappa accuracy sebesar 94.1%, minimum producer’s accuracy sebesar 90.1%, dan minimum user’s accuracy sebesar 88.6%. Peubah yang paling berpengaruh dalam pembangunan algoritma decision tree dalam deteksi deforestasi mangrove adalah ?BSI untuk peubah spektral dan substrat untuk peubah sosio-geo-biofisik. Penghalusan citra pasca klasifikasi mampu menghilangkan noise dalam hasil klasifikasi. Deforestasi di Kabupaten Langkat sebagian besar terjadi akibat alih fungsi mangrove menjadi tambak dan perkebunan.

This research presents the development of a machine learning-based method for detecting mangrove cover change. The study was conducted on deforestation events that occurred in Langkat Regency, Indonesia, from 2018 to 2024. A change map was generated using the image differencing method. The mangrove cover change model was constructed using a combination of variables derived from the image differencing results and a set of socio-geo-biophysical parameters. This study demonstrates that the best model was constructed using a combination of image differencing and socio-geo-biophysical variables with post-classification image smoothing, resulting in an overall accuracy of 95.1%, a kappa accuracy of 94.1%, a minimum producer’s accuracy of 90.1%, and a minimum user’s accuracy of 88.6%. The most influential variables in the development of the decision tree algorithm were the change in the Bare Soil Index (?BSI) as a spectral variable and substrate type as a socio-geo-biophysical variable. Post-classification image smoothing effectively eliminated noise from the classification results. Deforestation in Langkat Regency was found to be predominantly driven by the conversion of mangrove areas into aquaculture ponds and plantations.