Advanced Dynamic Monitoring and Precision Analysis of Soil Salinity in Cotton Fields Using CNN ‐Attention and UAV Multispectral Imaging Integration
Land Degradation & Development · 2025-06 · 36 (10) · 3472-3489
DOI: 10.1002/ldr.5578 ↗Kurz gefasst
Fachartikel in der Zeitschrift Land Degradation & Development, online erschienen am 14.05.2025. Verfasst von Jiao Tan und 8 weiteren Autor:innen. Die Arbeit wurde bisher 7-mal zitiert (Crossref) und nennt 65 Literaturangaben.
Abstract (Originalsprache)
Accurate and timely estimates of crop exposure to salt stress are essential for monitoring crop growth and implementing effective management practices. However, most contemporary research has focused on single‐period soil salinity estimations and relied on traditional machine learning methods, which struggle to account for the temporal dynamics of soil salinity. This study proposed a modeling framework that combined multi‐temporal UAV multispectral imagery and measured soil salinity data to estimate agricultural soil salinity. Key growth stages in soil preparation, squaring stage, flowering stage, and boll opening stage were evaluated and combined with field‐measured soil salinity values. Based on different combinations of inputs of indices, textures, and spectral reflectance, recursive feature cancelation cross‐validation (REFCV), Elastic Net, and XGBoost were used for selection of features extracted from multispectral imagery. The selected features were used to train and test random forest (RF) and Convolutional Neural Network‐Attention (CNN‐Attention) models. The results of the study show that (1) the REFCV algorithm is stable in feature selection, the EN algorithm is more prominent in the squaring stage and flowering stage, and XGBoost results are optimal. (2) After incorporating texture features, the model's R 2 showed varying degrees of improvement. the R 2 value of the RF model in Saline‐alkaline farmland increased to 0.912 and the RMSE decreased to 0.207, while in high standard farmland, the R 2 reached 0.891 and the RMSE decreased to 0.255, with a significant improvement in model accuracy. (3) Overall, the CNN‐Attention model demonstrated a higher prediction accuracy at all time points and feature combinations. (4) In different scenarios, the RF model is suitable for long‐term stable monitoring tasks, and the CNN‐Attention model has significant advantages in complex feature extraction and dynamic change capture.
Zitieren
APA
Tan, J., Ding, J., Li, J., Han, L., Cui, K., Li, Y., Wang, X., Hong, Y., & Zhang, Z. (2025). Advanced Dynamic Monitoring and Precision Analysis of Soil Salinity in Cotton Fields Using CNN ‐Attention and UAV Multispectral Imaging Integration. Land Degradation & Development, 36(10), 3472-3489. https://doi.org/10.1002/ldr.5578
Vancouver
Tan J, Ding J, Li J, Han L, Cui K, Li Y, et al. Advanced Dynamic Monitoring and Precision Analysis of Soil Salinity in Cotton Fields Using CNN ‐Attention and UAV Multispectral Imaging Integration. Land Degradation & Development. 2025;36(10):3472-3489. doi:10.1002/ldr.5578
BibTeX
@article{Tan2025,
author = {Tan, Jiao and Ding, Jianli and Li, Jiangtao and Han, Lijing and Cui, Kuangda and Li, Yongkang and Wang, Xiao and Hong, Yanhong and Zhang, Zhe},
title = {Advanced Dynamic Monitoring and Precision Analysis of Soil Salinity in Cotton Fields Using CNN ‐Attention and UAV Multispectral Imaging Integration},
journal = {Land Degradation & Development},
year = {2025},
volume = {36},
number = {10},
pages = {3472--3489},
doi = {10.1002/ldr.5578}
}
EUVIMED