Assessing machine learning techniques for detailing soil map in the semiarid tropical region

Abstract

The major problem detailing soil map in large tropical country such as Indonesia is high cost and time-consuming. The machine learning technique is one of DSM methodologies that explores spatial patterns to predict soil class and soil attribute. K-nearest neighbours (KNN), random forest (RF) and support vector machine (SVM) are popular for detailing soil map in temperate country, but it is still rare to be applied in a tropical country. This study aimed to asses three machine learning in updating soil map from 1:50,000 to 1:25,000 scale in the semiarid tropical region. The existing soil map was collated and then derived environmental covariates representing soil-forming factors from the digital elevation model. There were 72 training datasets were originating from polygon soil maps used as input for these machine learning to recognize the pattern and predict soil class map in Bikomi Utara Sub District, Timor Tengah Utara Regency, Indonesia. Overall accuracy and kappa coefficient by KNN for the best three predictive soil maps were 74-75% and 0.62-0.63, respectively; and followed by SVM, 71-73% and 0.58-0.60; and the last RF, 69-75% and 0.55-0.63. This research revealed that machine learning of the KNN is potentially for updating soil map in a tropical semiarid area.