Pemetaan Objek di Lahan Sawah dengan Random Forest dan Support Vector Machine Berbasis Data Drone

Abstract

Pemetaan objek pada sepetak lahan sawah dilakukan dengan memanfaatkan citra drone multispektral yang memiliki resolusi spasial tinggi (GSD = 6,85mm/piksel). Citra multispektral yang diakuisisi menggunakan wahana udara tanpa awak (Unmanned Aerial Vehicle/UAV) dapat dimanfaatkan untuk mengklasifikasikan berbagai objek di lahan sawah berdasarkan karakteristik spektralnya. Penelitian ini bertujuan membandingkan kinerja algoritma Random Forest (RF) dan Support Vector Machine (SVM) dalam pemetaan objek di lahan sawah berbasis citra multispektral UAV. Tahapan penelitian meliputi akuisisi data UAV, pengolahan citra secara fotogrametri, penyusunan data training dan data validasi, klasifikasi menggunakan algoritma RF dan SVM, serta evaluasi akurasi berdasarkan confusion matrix. Hasil klasifikasi membedakan objek ke dalam tujuh kelas, yaitu padi sehat, padi kerusakan ringan, padi kerusakan sedang, padi kerusakan berat, tanah, air, dan rumput. Evaluasi akurasi menunjukkan bahwa algoritma RF menghasilkan nilai Overall Accuracy sebesar 77,27% dan koefisien Kappa sebesar 72,42%, sedangkan algoritma SVM menghasilkan nilai Overall Accuracy sebesar 80,91% dan koefisien Kappa sebesar 76,56%. Perbandingan hasil klasifikasi menunjukkan tingkat kesesuaian spasial sebesar 83,22%, sedangkan 16,78% area menunjukkan perbedaan hasil klasifikasi. Secara keseluruhan, algoritma SVM memberikan kinerja yang lebih baik dibandingkan RF dalam pemetaan objek di lahan sawah pada lokasi penelitian.

Mapping objects within a rice field was conducted using multispectral drone imagery with a high spatial resolution (GSD = 6.85 mm/pixel). Multispectral imagery acquired using an Unmanned Aerial Vehicle (UAV) can be utilized to classify various objects in paddy fields based on their spectral characteristics. This study aimed to compare the performance of the Random Forest (RF) and Support Vector Machine (SVM) algorithms for object mapping in paddy fields using UAVbased multispectral imagery. The research workflow included UAV data acquisition, photogrammetric image processing, preparation of training and validation datasets, image classification using the RF and SVM algorithms, and accuracy assessment based on a confusion matrix. The classification results identified seven object classes: healthy rice, lightly damaged rice, moderately damaged rice, severely damaged rice, soil, water, and grass. The accuracy assessment showed that the RF algorithm achieved an Overall Accuracy of 77,27% with a Kappa coefficient of 72,42%, while the SVM algorithm achieved an Overall Accuracy of 80,91% with a Kappa coefficient of 76,56%. The comparison of the classification results indicated a spatial agreement of 83,22%, while 16,78% of the study area showed differences between the two classification methods. Overall, the SVM algorithm outperformed the RF algorithm in mapping objects in the paddy field study area.