Evaluating UAV multispectral imagery, machine learning and image analysis techniques for mapping taro and sweet potato in a smallholder cropland in Swayimane, South Africa

Abstract

Accurate mapping of neglected and underutilised crop species (NUS) in complex smallholder fields is critical for enhancing foodsecurity, yet it remains a significant challenge due to landscape heterogeneity, spectral similarities between crops and the spatialcomplexity of strip intercropping systems. This study evaluates the efficacy of unmanned aerial vehicle (UAV) multispectralimagery for mapping taro and sweet potato, addressing the limitations of traditional pixel-based image analysis (PBIA), whichoften fails in fragmented agricultural landscapes. We quantitatively compare the performance of PBIA and object-based imageanalysis (OBIA) frameworks using a Gradient Tree Boost (GTB) classifier and determine the optimal input dataset by testing rawspectral bands, vegetation indices (VIs) and their combination. The findings demonstrate the clear superiority of the OBIA–GTBapproach: using the combined dataset, OBIA–GTB achieved a mean overall accuracy of 95% ± 3.66%, kappa = 0.9419. In contrast,PBIA- GTB yielded 83% ± 7.10% overall accuracy, with kappa = 0.8016. This study validates a robust, scalable workflow for NUSmapping, concluding that the integration of high-resolution UAV data, OBIA and GTB provides a precise solution for agriculturalmonitoring in heterogeneous, intercropped smallholder systems.