Seleksi Hyperspectral Band Menggunakan Recursive Feature Elimination untuk Prediksi Produksi Padi dengan Support Vector Regression
Abstract
Hyperspectral is a new technology in remote sensing that exploits hundreds of bands. Pusat Teknologi Inventarisasi Sumber Daya Alam Badan Pengkajian dan Penerapan Teknologi (PTISDA BPPT) applies hyperspectral in agriculture for yearly yield prediction. Hyperspectral images consist of large number of bands that require analysis to select features. One approach to reduce computational cost is to eliminate bands that do not add value to the regression and analysis method to be applied. In this research, we use a Recursive Feature Elimination (RFE) algorithm that is tailored to operate with Support Vector Regression (SVR) to perform band selection and regression simultaneously. Annual yield of paddy has been predicted with hyperspectral data using Support Vector Regression (SVR) algorithm. Regions used are Indramayu and Subang, and the altitude of the spectral acquisition is 2000 m (Hymap). This data is owned by PTISDA BPPT. RFE-SVR used in hyperspectral data was able to reduce about 30% of the bands, resulting in 70 bands out of 109 original bands with an RMSE value of 0.0901 and R2 value of 0.9874. Radial Basis Function (RBF) is the best kernel used in RFE-SVR having an RMSE value less than those of other kernels tested.
Collections
- UT - Computer Science [2254]