# CSCA-YOLOv8: A lightweight network model for evaluating drought resistance in mung bean

## TL;DR

This paper introduces a lightweight model for identifying drought resistance in mung beans using chlorophyll fluorescence images, enabling efficient deployment on low-power devices.

## Contribution

The novel CSCA-YOLOv8 model integrates a lightweight backbone and attention mechanisms to improve drought resistance identification in mung beans.

## Key findings

- The model reduces parameters by 24% and floating point operations by 35% compared to YOLOv8s.
- Accuracy improved by 2.52%, suitable for deployment on edge devices.
- The Mungbean Drought Dataset (MDD) was created with 4808 labeled images for training and validation.

## Abstract

Drought is one of the main factors affecting mung bean production in China. Screening drought-resistant germplasm resources and cultivating drought-resistant varieties are of great significance to the development of the mung bean industry in China. Combined with chlorophyll fluorescence imaging technology, this paper proposes a lightweight mung bean drought resistance identification network model based on YOLOv8, referred to as CSCA-YOLOv8. The model uses StarNet to replace the backbone network of YOLOv8 to reduce the size of the model. The C2f_Star module is introduced in the neck structure instead of the original C2f module. Then, in order to enhance the network’s attention to the key regions in the feature map, the Context Anchor Attention Mechanism (CAA) module is also introduced into the fourth C2f_Star module. Then, a CGBD module is proposed in the neck structure to reconstruct the ordinary convolution to improve the feature extraction ability of the model for small targets. Finally, the SIoU loss function is used to replace CIoU to accelerate the convergence of the model. In the actual data analysis, we used the collected 4808 chlorophyll fluorescence images of the natural mung bean population under drought stress to make the Mungbean Drought Datatset(MDD) and made classification labels for each image according to different drought resistance levels, which were 0, 1, 2, 3, 4 and 5. We also verified the excellent performance and generalization performance of the model using the collected MDD dataset. The final experimental results show that compared with the YOLOv8s baseline model, the number of parameters of our proposed algorithm is reduced by 24%, the floating point number is reduced by 35%, and the accuracy is improved by 2.52%, which supports the deployment on embedded edge devices with limited computing power. Therefore, our proposed algorithm has great potential in the field of drought resistance identification and germplasm selection of mung bean.