A Multi-Task Deep Learning Framework for In-Field Leafy Pumpkin (Telfairia occidentalis) Classification

Abstract

Telfaria occidentalis Hook f. is native to West Tropical Africa. The accurate in-field identification and disease assessment of this leafy pumpkin are critical for optimizing crop yield and minimizing chemical interventions in precision agriculture. However, conventional computer vision approaches frequently underperform in real-world agronomic environments due to cluttered background vegetation, severe leaf occlusions, and erratic ambient lighting. Furthermore, extant literature predominantly treats crop identification, leaf segmentation, and disease diagnosis as isolated tasks, introducing substantial computational latency that limits operational efficiency on resource-constrained deployment systems. To address these limitations, this study introduces a unified, multi-task deep learning framework utilizing the Ultralytics architecture to simultaneously achieve in-field pumpkin foliage identification, instance-level semantic segmentation, and automated disease severity estimation. By leveraging a single-stage, multi-head neural network topology, the framework isolates individual, overlapping pumpkin leaves from weed-infested backgrounds while concurrently segmenting localized necrotic lesions caused by prevalent foliar pathologies. Disease severity is computationally quantified in real time by computing the pixel-wise ratio of detected lesion masks to the localized parent leaf segment mask. Experimental evaluations demonstrate that this joint multi-task pipeline not only matches or exceeds the precision metrics of isolated, single-task architectures but also significantly reduces computational latency and inference overhead. This system provides a scalable, end-to-end machine vision solution suitable for deployment on mobile apps and autonomous agricultural machinery, successfully bridging the gap between high-throughput plant phenotyping and real-time field diagnostics.