Hyperspectral Estimation of Chlorophyll Content in Ginseng Fruit Leaves Based on Wavelet Transform and VCPA-GA Algorithm
Leaf Chlorophyll Content (LCC) is vital for both direct and indirect plant growth and development. Accurate monitoring of LCC in ginseng fruits provides essential data for assessing their photosynthetic and nutritional status, which is beneficial for the development of precision agriculture. Traditi...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Committee of Tropical Geography
2025-03-01
|
| Series: | Redai dili |
| Subjects: | |
| Online Access: | https://www.rddl.com.cn/CN/10.13284/j.cnki.rddl.20240280 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Leaf Chlorophyll Content (LCC) is vital for both direct and indirect plant growth and development. Accurate monitoring of LCC in ginseng fruits provides essential data for assessing their photosynthetic and nutritional status, which is beneficial for the development of precision agriculture. Traditional chemical analyses in laboratories require a large number of samples, which are not only time-consuming and destructive, but also fail to meet the precise management needs of extensive fields. Although some handheld devices can measure the leaf LCC accurately and quickly without causing damage, they cannot provide large-scale information. Hyperspectral remote sensing is widely applied for rapid and non-destructive LCC monitoring because of its strong continuity and abundant data. In this study, we used ginseng fruit leaf hyperspectral data and the corresponding LCCs as datasets. We applied the Discrete Wavelet Transform (DWT) to extract the low-frequency coefficients from the 0-10 layers of the hyperspectral data. We then conducted a Pearson correlation analysis on the 0-10 layer spectral datasets and their corresponding LCCs. We combined Variable Combination Pattern Analysis (VCPA) with Genetic Algorithm (GA), employing the combined VCPA-GA algorithm to extract sensitive bands from the full spectrum and each decomposed layer of the ginseng fruit leaf. Finally, we established estimation models for the ginseng fruit LCC using the Back Propagation Neural Network (BPNN), GA-BPNN, Particle Swarm Optimization (PSO)-BPNN, and BP-AdaBoost neural network models. Among the four machine-learning models, the BP-AdaBoost neural network exhibited the best overall predictive performance. The predictive performance of the PSO-BPNN model was similar to that of the BPNN model, whereas the GA-BPNN model exhibited the lowest predictive performance. This study shows: (1) The 1-5 layer DWT spectra accurately reflect the overall characteristics of the original spectrum, with a decrease in correlation at each layer beyond the fifth layer, and the spectra beyond the seventh layer no longer represent the overall features of the original spectrum. This is because the wavelet transform process has some errors that increase with the number of decomposition layers. (2) The VCPA-GA hybrid variable selection algorithm merges the strengths of the VCPA and GA, addressing the tendency of the VCPA to select fewer variables and overcoming GA's limitations in handling many variables which can lead to overfitting, providing a theoretical basis for estimating ginseng fruit LCC using hyperspectral remote sensing. (3) Among the four machine-learning models, predictions from to 1-2 and 6-7 layers were generally lower than those of the 0 layer, while predictions from the 3–5 layers are higher, showing an overall trend of initial increase followed by a decrease as the number of wavelet decomposition layers increased. (4) Ginseng fruit leaf hyperspectral data processed by the DWT-VCPA-GA algorithm with a 4-layer DWT spectrum yielded the best predictive performance in the BP-AdaBoost regression model, with R2=0.919, mean absolute percentage error = 2.090%, and relative percentage difference = 3.900. (5) After optimizing the BPNN regression model with various algorithms, only some optimized models improved their predictive performance and accuracy to a certain extent, making the choice of the right optimization algorithm crucial for model improvement. |
|---|---|
| ISSN: | 1001-5221 |