Automated detection of individual clove trees for yield quantification in northeastern Madagascar based on multi-spectral satellite data

Abstract

There is an increasing demand for clove products, mainly dried buds and essential oil on global markets. Consequently, the importance of clove trees as a provisioning service is increasing at the local level, particularly for smallholders cultivating clove trees as cash crops. Due to limited availability of data on local production, using remote sensing-based methods to quantify today's clove production is of key interest. We estimated the clove bud yield in a study site in northeastern Madagascar by detecting individual clove trees and determining relevant productio n systems, including pasture and clove, clove plantation and agroforestry systems. We implemented an individual tree detection method based on two machine learning approaches. Specifically, we proposed using a circular Hough transform (CHT) for the automated detection of individual clove trees. Subsequently, we implemented a tree species classification method using a random forests (RF) classifier based on a set of features extracted for relevant trees in the above production systems. Finally, we classified and mapped different production systems. Based on the number of detected clove trees growing in a clove production system, we estimated the production system-dependent clove bud yield. Our results show that 97.9% of all reference clove trees were detected using a CHT. Classifying clove and non-clove trees resulted in a producer accuracy of 70.7% and a user accuracy of 59.2% for clove trees. The classification of the clove production systems resulted in an overall accuracy of 77.9%. By averaging different clove tree yield estimates obtained from the literature, we estimated an average total yield of approximately 575 tons/year for our 25,600 ha study area. With this approach, we demonstrate a first step towards large-scale clove bud yield estimation using remote sensing data and methodologies.