Towards Resource-Efficient AI for Opuntia Species Identification and Structural Characterization

Opuntia (prickly pear cactus) is an economically and ecologically important crop across arid and semi-arid regions of the Global South, yet species identification and structural monitoring (cladode and fruit counts) remain manual, slow, and expert-dependent. Most prior computer vision approaches to crop monitoring assume access to UAV platforms, multispectral sensors, or large curated datasets resources frequently unavailable in low-resource agricultural settings. This work investigates whether accessible, low-cost tools a standard consumer camera and lightweight, freely available object detection models can deliver usable accuracy for Opuntia monitoring without specialized equipment. Using 429 field-collected RGB images spanning seven Opuntia species, we trained and compared three detection architectures (YOLOv5, YOLOv8, and Roboflow's Object Detection Model 3) via transfer learning and combined the two strongest models in a confidence-weighted ensemble. YOLOv8 substantially outperformed YOLOv5 (81.3% vs. 31.3% mAP50), and the ensemble raised recall to 97.5% (vs. 83.3% for the best single model) at a measurable precision cost a trade-off relevant to different field deployment needs. The main source of error was confusion between two morphologically similar species, reflecting a known limitation of RGB-only imagery for fine-grained taxonomic classification. These results suggest that meaningful crop-monitoring accuracy is achievable using equipment already accessible to students, farmers, and practitioners in resource-constrained settings, offering a starting point for more accessible precision-agriculture tools in the Global South.