A new feature selection algorithm combining genetic algorithm, exponential decay function, and machine learning to realize hyperspectral estimation of winter wheat leaf area index

•A new band screening algorithm named CRDGA was proposed.•CRDGA had lower fitness value than GA.•The number of bands screened by the CRDGA was small than GA.•CRDGA can control the number of bands more flexibly than GA.•KNNR was more suitable as the fitness evaluation function of the new algorithm. T...

Full description

Saved in:
Bibliographic Details
Published in:Computers and electronics in agriculture Vol. 230; p. 109851
Main Authors: Yang, Chenbo, Bai, Juan, Sun, Hui, Bi, Rutian, Song, Lifang, Muhammad, Amjad, Wang, Chao, Zhao, Yu, Yang, Wude, Xiao, Lujie, Zhang, Meijun, Song, Xiaoyan, Feng, Meichen
Format: Journal Article
Language:English
Published: Elsevier B.V 01.03.2025
Subjects:
agriculture
algorithms
Competitive adaptive reweighted sampling
Continuous reweighted decay genetic algorithm
electronics
Genetic algorithm
Hyperspectral
hyperspectral imagery
Leaf area index
Machine learning
normal distribution
regression analysis
winter wheat
Continuous reweighted decay genetic algorithm
Competitive adaptive reweighted sampling
Genetic algorithm
Hyperspectral
Machine learning
Leaf area index
ISSN:0168-1699
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•A new band screening algorithm named CRDGA was proposed.•CRDGA had lower fitness value than GA.•The number of bands screened by the CRDGA was small than GA.•CRDGA can control the number of bands more flexibly than GA.•KNNR was more suitable as the fitness evaluation function of the new algorithm. The leaf area index (LAI) is an important parameter that can reflect the growth status of the winter wheat population. Using hyperspectral technology for rapid and non-destructive estimation of LAI can provide some reference for adjusting field management measures to promote healthy growth of winter wheat. Screening hyperspectral information is an important direction in hyperspectral research. Genetic algorithm (GA) is a commonly used variable selection method. However, traditional GA cannot effectively control the results of variable screening. Meanwhile, it is expected to further improve the variable screening effect by using more machine learning algorithms as fitness evaluation functions. In this regard, this study combined the exponential decay function in competitive adaptive reweighted sampling algorithm (CARS) and the GA, and combined multiple machine learning algorithms as fitness evaluation functions to propose the continuous reweighted decay genetic algorithm (CRDGA), to better screen the important bands of winter wheat LAI. Finally, hyperspectral estimation models for winter wheat LAI were constructed using multiple machine learning algorithms. The main results were as follows: Compared to the number of bands screened by CARS and GA, the number of bands screened by CRDGA was fewer, with 44, 32, 24, 10, 7, 19, and 197 bands, respectively. Using partial least squares regression, support vector regression (SVR), K-nearest neighbor regression (KNNR), and gaussian process regression (GPR) as fitness evaluation functions all had a faster convergence rate, other fitness evaluation functions all had a slower convergence rate. However, only SVR, KNNR, and GPR as fitness evaluation functions had fast running rates in both GA and CRDGA. When using the same fitness evaluation function to run GA and CRDGA to achieve the optimal fitness value, the fitness value obtained by the CRDGA algorithm was lower than that obtained by the GA algorithm. Among all the constructed models, KNNR was used to construct the model with the highest accuracy after the bands were screened by CRDGA-KNNR, and its R2c, RMSEc, R2v, and RMSEv were 0.6511, 2.0560, 0.7403, and 1.8004, respectively, which could realize the hyperspectral estimation of LAI in winter wheat. Overall, the CRDGA algorithm proposed in this study performs better than CARS and GA algorithms.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:0168-1699
DOI:10.1016/j.compag.2024.109851