Knowledge Gaps and Experimental Framework for GreenEdge-AgriAI: Energy-Aware Plant Disease Detection

Plant pathogens and pests are estimated to destroy a substantial share of global crop yield each year, and pressure on food security will intensify as population and climate stress both grow [1][2]. Smartphone penetration in smallholder-farming regions of South Asia and Africa has grown rapidly over the same period, creating an opportunity for camera-based, on-device disease diagnosis that does not depend on network connectivity [3]. Since Mohanty et al. [4] first demonstrated deep-learning-based leaf-disease classification, a large body of lightweight-architecture research has pursued high accuracy within tight parameter and latency budgets, but almost none of it treats the energy consumed by the target device as a first-class design objective [5][6].

Conference 1 of this SGS Initiative programme reviewed more than twenty-five such systems, synthesised a five-dimensional taxonomy — architecture design, model compression, training strategy, edge deployment and explainability — and proposed TinyCNN-Lite as a candidate architecture with a projected performance target. That review concluded that no existing system satisfies all five dimensions simultaneously, and that energy per inference in particular has been left undefined as a design constraint.

This Conference 2 paper carries the programme into its mandated second phase. It (i) restates the Conference 1 observations as five precisely defined knowledge gaps (Section 2), (ii) specifies the complete GreenEdge-AgriAI methodological and experimental framework built to close them, including hypotheses, dataset protocols, hardware-profiling methodology and a formal Green AI metric suite (Section 3), and (iii) reports preliminary design-stage validation confirming that the framework is architecturally viable before full experimental evaluation at Conference 3 (Sections 3.6 and 4). Section 5 concludes with the experimental roadmap for Conference 3.