Improved MobileVit deep learning algorithm based on thermal images to identify the water state in cotton

文献类型： 外文期刊

第一作者： Jin, Kaijun

作者： Jin, Kaijun;Zhang, Jihong;Liu, Ningning;Li, Miao;Ma, Zhanli;Wang, Zhenhua;Zhang, Jinzhu;Yin, Feihu;Jin, Kaijun;Zhang, Jihong;Liu, Ningning;Li, Miao;Ma, Zhanli;Wang, Zhenhua;Zhang, Jinzhu;Jin, Kaijun;Zhang, Jihong;Liu, Ningning;Li, Miao;Ma, Zhanli;Wang, Zhenhua;Zhang, Jinzhu;Jin, Kaijun;Zhang, Jihong;Liu, Ningning;Li, Miao;Ma, Zhanli;Wang, Zhenhua;Zhang, Jinzhu;Yin, Feihu

作者机构：

关键词： Cotton; Water state identification; Computer vision; Improved MobileVit algorithm; Deep learning

期刊名称：AGRICULTURAL WATER MANAGEMENT （ 影响因子：6.5； 五年影响因子：6.9 ）

ISSN： 0378-3774

年卷期： 2025 年 310 卷

页码：

收录情况： SCI

摘要： Thermal imaging combined with deep learning algorithms offers an efficient and non-invasive method for monitoring crop water status, facilitating precise irrigation management over large agricultural areas. This study introduces a method for identifying the moisture state of cotton using an enhanced MobileVit deep learning algorithm. This approach incorporates the Efficient Channel Attention (ECA) mechanism into the Fusion component of the MobileVit model, optimizes the first convolution in the Fusion component by replacing it with Depthwise Separable Convolution (DsConv), and substitutes the Local representation with the MobileOne block. These enhancements aim to improve model performance while maintaining its compact size. A dataset of thermal images of cotton canopies representing three different water states was developed for this study. Ablation studies were performed to evaluate the effect of each modification. Grad-CAM was utilized to illustrate the final layer features of the proposed algorithm. Various deep learning models were also trained, tested, and validated, allowing for a comparative analysis of the proposed model against traditional deep learning models in identifying cotton moisture states. The results show that the F1-score of the proposed model reaches 0.9677, achieving a recognition speed of 50.370 ms while maintaining a size of 4.94 M, outperforming other classical deep learning models. The findings of this study provide technical support for the development of future precision irrigation systems. The relevant code and datasets will be made available on GitHub (https://github.com/kingcuzamu/i dentifying-cotton-water-state) upon publication.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序