NAS-FertiSense: A Lightweight Crop Fertilizer Deficiency Monitoring System Framework Based on Dual Modality

Agricultural fertilizer deficiency monitoring is crucial for precision farming, yet current approaches face significant limitations. High deployment costs associated with deep learning solutions and the subjectivity of manual experience-based methods hinder their practicality. To address these challenges, this paper introduces NAS-FertiSense, an embedded-friendly vision-sensor dual-modal fusion framework. It employs an improved NCC algorithm for rapid leaf localization and static data acquisition, followed by a three-level collaborative verification strategy: (1) visual modality using an adaptive threshold AGAST algorithm to extract leaf deficiency features (e.g., yellowing patch density, edge breakpoints), (2) sensor modality providing quantitative NPK data based on dynamically compensated soil values, and (3) sensitivity-adjusted retests with HSV color space arbitration for conflict resolution. Additionally, it utilizes a spatiotemporal alignment mechanism involving 10 samplings per planting unit and confidence-weighted voting for the final diagnosis. Experiments on a self-built dataset of four crops (including corn and tomatoes) demonstrate a 93.7% recognition accuracy, comparable to lightweight neural networks, while the algorithm volume is only 1/10, significantly reducing hardware resource demands for low-power embedded field deployment. When implemented on the OpenMV+STM32 platform, the system completes detection within a fixed cycle time at an average power level, enabling a low-cost, plug-and-play monitoring solution that requires no labeled data.