Land Suitability Modeling For Paddy Field With A Machine Learning Approach Using Physical Variables Budi Siswanto (a*), Ketut Wikantika (b), Albertus Deliar (b), Tri Muji Susantoro (c)
a) Doctoral Study Program of Geodesy and Geomatics Engineering, Faculty of Earth Science and Technology, Institut Teknologi Bandung, Jl. Ganesha No. 10, Bandung 40132, Indonesia
b) Remote Sensing and Geographic Information Science Research Group, Department of Geodesy and Geomatics, Faculty of Earth Sciences and Technology, Institut Teknologi Bandung, Bandung 40132, Indonesia
c) Research Center for Remote Sensing-The Research Organization for Aeronautics and Space, Kompleks Cibinong Science Center-BRIN, Jalan Raya Bogor KM. 46, Cibinong 16911, Jawa Barat, Indonesia
Abstract
As the human population continues to grow, the demand for staple food increases. At the same time, paddy fields, which are a key factor influencing food supply, are decreasing. This condition can lead to a food availability crisis, thus increasing the need to expand agricultural land, in particular paddy fields. Consequently, assessing land suitability for paddy fields across various locations becomes necessary.
In this research, the assessment of land suitability for paddy fields in Majalengka and Indramayu Regency was conducted using a machine learning approach. The algorithms employed included Random Forest (RF), Logistic Regression (LR), and Geographically Weighted Logistic Regression (GWLR). The independent variables used were elevation, land inclination, rainfall, distance from streams, topographic wetness index (TWI), soil pH, soil texture, rock formation, and soil permeability. The dependent variable was paddy field land cover.
This research demonstrated that the Geographically Weighted Regression (GWR), Random Forest (RF), and Logistic Regression (LR) models achieved accuracies of 63.48%, 62.86%, and 61.84%, respectively, in classifying paddy field and non-paddy field areas. Based on these results, it can be concluded that GWR achieved the highest accuracy. In the GWR model, a Fixed Gaussian Kernel Weighting Function was used, with a deviance value of 330 and a Percent Deviance Explained of 25%.
Keywords: Land suitability, Machine learning, Random Forest, Logistic Regression, Geographically Weighted Logistic Regression
Topic: Interdisciplinary Earth Science and Technology
