Developing decision-tree machine learning algorithm for identifying cocoa agroforestry: a case study in North Luwu Regency, Indonesia

This paper was not presented at the conference.

Abstract

This paper was not presented at the conference.

This paper describes the development of a decision-tree machine learning algorithm for detecting and identifying monoculture cocoa and agroforestry cocoa plantations in Luwu Utara Regency, South Sulawesi Province of Indonesia. The main objective of this study was to determine the best decision tree algorithm for identifying cocoa plants in agroforestry patterns versus monocultures. Because most cocoa plants are grown by smallholder farmers on a small scale, it is challenging to find large areas of cocoa plants. Finding cocoa plants in small spaces, whether planted in monoculture or under forest stands using an agroforestry system, is challenging. In this study, the author developed a cocoa plant detection algorithm by considering aspects (variables) related to the tradition of choosing plant sites according to the geo-socio-biophysical such as elevation, slope, distance from settlements, roads, rivers, cocoa suitability class, and land use land cover, combined with spectral characteristics of cocoa plants, i.e., some vegetation indices, as well as bare land and wetness indices derived from Landsat TM and SPOT-7 imageries. To measure the weight of each variable in the decision tree, the study applies the "Brute Force" approach using the entropy derived using the information gain, gini index, and gain ratio method.  The study found that the most significant variables for identifying agroforestry and monoculture cocoa using TM images are NDVI, NBR, NDWIG, VDI, DEM, distance from settlements, and land use, with an overall accuracy of 83%. The SPOT-7 imagery gave more accurate detection, having approximately 93% overall accuracy, derived using NDVI, NDWIG, NRGI, DEM, and distance from settlements, roads, rivers, and land use.