Developing a decision tree machine learning algorithm for coffee agroforestry detection using SPOT-7 satellite imagery in Tanggamus District

Abstract

This study describes the development of a machine learning algorithm to detect and identify agroforestry coffee plants using a machine learning approach. The primary data used in this study are SPOT-7 satellite image, which are then transformed into synthetic images of NDVI, VDVI, VARI, and NRGI. To improve accuracy, this study also integrates synthetic images with socio-geo-biophysical variables, including road proximity, river proximity, settlement proximity, elevation, slope, and visual land cover. This study found that the best algorithm for detecting agroforestry and monoculture coffee plants utilizes a decision tree algorithm with information gain parameters and variables, including NDVI, VDVI, NRGI, elevation, road proximity, river proximity, and land cover. This algorithm utilizes a maximum depth of 31, without pruning or pre-pruning, a minimum leaf size of 51, a minimum size for splitting of 70, and a pre-pruning alternative of 60. This algorithm produces an overall accuracy of 94.55% and a kappa accuracy of 93.7%. The most influential variable in the detection model was land cover. Agroforestry coffee had a user accuracy (precision) of 78.8% and a producer accuracy (recall) of 80.81%. The machine learning algorithm model was able to accurately detect and identify the distribution of agroforestry and monoculture coffee.