DROUGHT RECOGNITION FOR THE SOYBEAN PLANT BASED ON LIGHTWEIGHT DEEP LEARNING MODEL

Abstract

To address the issues of low soybean self-sufficiency and drought - related production constraints, multispectral imaging can non - invasively detect crop characteristics, while deep convolutional networks identify drought from multispectral images. However, large models face mobility limitations due to high computational requirements.This study presents a mobile detection approach integrating ReliefF feature screening and lightweight convolutional neural networks, and develops the "Early Acknowledgment for Soybean Drought" App. It embeds a lightweight model that optimizes 37 - dimensional soybean canopy multispectral features via ReliefF and uses a three - layer 1D convolutional network for drought identification.The model achieves 96.88% classification accuracy on the self - built dataset, with an inference time of 18 ms, a size under 30 MB, and less than 60 MB memory usage on mobile. The APP integrates the multispectral camera SDK and PyTorch inference engine, enabling real - time spectral analysis. Field tests show its one - button operation, low learning curve for farmers, and significant water - saving and yield - increasing effects, offering a lightweight, high - precision mobile solution for soybean drought management and promoting smart agriculture development.

Plant phenotyping relevance

マルチスペクトル画像からダイズの干ばつ状態を推定する軽量深層学習モデルとモバイルアプリを開発しており、植物状態の取得・推定手法が研究の中心である。

Code and data availability

The paper describes a self-built soybean canopy multispectral dataset (Parrot Sequoia camera, 32 training samples) and a Conv-ReliefF model, but no blocks contain any data or code availability statement, public repository, DOI deposit, or authors' URL for the dataset, images, trained model, or app code. Only the paper-

No evidence-backed public reproduction asset is currently recorded.