SALT LAND CLASSIFICATION BASED ON PANSHARPENING IMAGES USING RANDOM FOREST ALGORITHM (CASE STUDY: PATI DISTRICK)

• Main Author: Diastarini
• Format: Theses
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/74433
• Institution: Institut Teknologi Bandung

Researchers have previously used machine learning to classify Salt Land based on medium-resolution satellite imagery. However, the analysis and policymaking of national salt production require more detailed information in greater spatial object resolution. This study proposes classifying Salt Land into Active and Inactive based on image pansharpening using the Random Forest algorithm. Pansharpening was done on Pleiades and Sentinel 2 satellite imagery to produce new images with high spatial and spectral resolution. Pansharpening method used are High Pass Filter (HPF), Nearest Neighbor Diffusion Pansharpening (NNDiffuse), À Trous Algorithm-Based Wavelet Transform (AWT), and University of New Brunswick (UNB) PanSharp. The result shows that the best pansharpening method is UNB PanSharp, resulting in an RMSE value of 7.1. Furthermore, 73 variables were derived from pansharpening images and National Digital Elevation Model (DEMNAS) to create a Random Forest model. Analysis was then carried out to determine the level of variable importance and the best Salt Land classification scheme. The result shows: 1) The ranking of the variable group contributions in building a Random Forest classification model is texture index > spectral variable > red-edge index > topographical variable > vegetation and building index > biophysical parameters > water index; 2) scheme 9 is the most Efficient Random Forest classification model. This model used 37 most important variables and 100 decision trees, achieving an OA of 66.699% and a Kappa coefficient of 0.394. This study can provide a good reference for developing Salt Land mapping and policies for practitioners or stakeholders.