# Selection of Optimal Diagnostic Positions for Early Nutrient Deficiency in Cucumber Leaves Based on Spatial Distribution of Raman Spectra

## TL;DR

This paper introduces a method to identify optimal spots on cucumber leaves for early detection of nutrient deficiencies using Raman spectroscopy, improving accuracy in precision agriculture.

## Contribution

The study proposes a novel method to select optimal diagnostic positions on cucumber leaves for early nutrient deficiency detection using Raman spectroscopy.

## Key findings

- Spectral similarity is less stable during early leaf development or under short-term nutrient stress.
- Low-similarity data is predominantly found at leaf margins, near main veins, and at the leaf base.
- Excluding low-similarity data improves diagnostic model performance with higher precision, recall, and F1 scores.

## Abstract

Accurate diagnosis of crop nutritional status is critical for optimizing yield and quality in modern agriculture. This study enhances the accuracy of Raman spectroscopy-based nutrient diagnosis, improving its application in precision agriculture. We propose a method to identify optimal diagnostic positions on cucumber leaves for early detection of nitrogen (N), phosphorus (P), and potassium (K) deficiencies, thereby providing a robust scientific basis for high-throughput phenotyping using Raman spectroscopy (RS). Using a dot-matrix approach, we collected RS data across different leaf positions and explored the selection of diagnostic positions through spectral cosine similarity analysis. These results provide critical insights for developing rapid, non-destructive methods for nutrient stress monitoring in crops. Results show that spectral similarity across positions exhibits higher instability during the early developmental stages of leaves or under short-term (24 h) nutrient stress, with significant differences in the stability of spectral data among treatment groups. However, visual analysis of the spatial distribution of positions with lower similarity values reveals consistent spectral similarity distribution patterns across different treatment groups, with the lower similarity values predominantly observed at the leaf margins, near the main veins, and at the leaf base. Excluding low-similarity data significantly improved model performance for early (24 h) nutrient deficiency diagnosis, resulting in higher precision, recall, and F1 scores. Based on these results, the efficacy of the proposed method for selecting diagnostic positions has been validated. It is recommended to avoid collecting RS data from areas near the leaf margins, main veins, and the leaf base when diagnosing early nutrient deficiencies in plants to enhance diagnostic accuracy.