Decision-tree-based machine learning for detecting coffee agroforestry using SPOT-7

Abstract

This study develops a decision-tree-based machine-learning (ML) approach to identify coffee agroforestry plants using SPOT-7 satellite imagery. The algorithm was developed by examining the combination of image indices derived from SPOT-7 and biophysical variables. Detection using spectral variables is often hampered by spectral similarity between vegetation cover classes. This study found that a ML method that combines spectral and biophysical variables can significantly improve overall accuracy, from 60.4% (using conventional spectral variables alone) to 94% (using integrated spectral-biophysical variables). For detecting and identifying agroforestry coffee classes typically found under tree canopies, the addition of the “land cover” variable published by the Ministry of Environment and Forestry contributes significantly to the classification of agroforestry coffee. Important variables identified in this model are normalized difference vegetation index (NDVI), visible difference vegetation index (VDVI), normalized red-green vegetation index (NRGI), elevation, and land cover.