Assessment of machine learning classifiers in mapping the cocoa-forest mosaic landscape of Ghana

The absence of clear-cut directives on the optimal classifier for land-use land-cover (LULC) classification of Ghana's cocoa landscape presents a practice gap. This has resulted in monitoring challenges since it is difficult to effectively compare land cover maps because they differ in the classifiers. In this paper, we explored the performance of four commonly used machine learning classifiers in the cocoa landscape to accurately determine the option that best segregates the different vegetation classes. Specifically, the accuracy with which k-Nearest Neighbors (kNN), Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forest (RF) classifiers mapped the cocoa landscape in Juaboso, Ghana, was compared. A pre-processed Sentinel-2 image, 352 training and 151 validation points collected through field Global Positioning System survey were used. The image was classified using the kNN, ANN, SVM and RF classifiers. The accuracies of the LULC maps were assessed using overall accuracy (OA), Cohens’ kappa (k). Also, practitioners with practical knowledge of the land cover classes and their distribution in the landscape subjected the map to visual inspection. The OA and k values indicated RF (OA=84.77%, k = 0.801), kNN (OA = 84.11%, k = 0.796), ANN (OA = 76.13%, k = 0.7), and SVM (OA = 81.45%, k = 0.762) all performed well in classifying the landscape with a satisfactory agreement. Additionally, there are no clear-cut classifier that experts in remote sensing should apply while mapping Ghana's cocoa environment. At any point, using the classifier that most accurately represents the landscape is crucial and should be prioritized. Therefore, guidance on the choice of classification algorithms by researchers and practitioners for mapping the cocoa landscape of Ghana must not be limited to the overall accuracies and kappa only. Instead, operationalising a mapping and validation framework that incorporates experts’ review will yield a LULC map that better represents the landscape.