Detection of water content and size of peas based on hyperspectral imaging combined with 2D-CNN and irregular polygon size measurement techniques

Abstract

Pea storage stability and germination rely on moisture content and morphology, but traditional destructive methods cause sample damage, low efficiency, and subjective errors, limiting practical use. To overcome the destructive and inefficient limitations of traditional methods for pea quality assessment, this study develops an integrated, non-destructive framework for the simultaneous and rapid measurement of pea moisture content and size using hyperspectral imaging combined with deep learning. We innovatively converted one-dimensional spectral data into two-dimensional texture images via Gramian Angular Field (GAF) encoding and input them into a residual 2D Convolutional Neural Network (2D-CNN) for moisture prediction. For dimensional analysis, a novel algorithm based on irregular polygon geometry was proposed to accurately measure pea length and width. The GAF-2D-CNN model achieved superior performance for moisture prediction (prediction set R²=0.9818, RMSEP=0.0318 %, RPD=7.4780), significantly outperforming 1D-CNN, Least Squares Support Vector Machine (LSSVM), and Partial Least Squares Regression (PLSR) models. The dimensional algorithm also demonstrated high accuracy, especially for length measurement (R²=0.9946, RPD=13.94). This framework provides a robust, accurate, and high-throughput solution for automated pea quality grading, offering significant potential for applications in precision agriculture and storage management.

Plant phenotyping relevance

ハイパースペクトル画像、深層学習、形状アルゴリズムを統合し、エンドウの水分含量とサイズという植物形質を非破壊・高スループットに推定する方法の開発が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.