What question did this study set out to answer?

This research aims to develop a model for precise wheat head detection in challenging agricultural conditions.

March 26, 2026Open Access

Bidirectional state space modeling for lightweight and robust wheat head detection in complex agricultural environments

Key Points

This research aims to develop a model for precise wheat head detection in challenging agricultural conditions.
Proposed the Mamba-WheatNet model that integrates selective state space modeling and CNN
Utilized Bidirectional Selective State Space Module (BSSMBlock) for enhanced spatial operations
Implemented Residual Depthwise Separable Block (RDSBlock) for improved feature extraction
Conducted ablation studies and comparative experiments using the GWHD-2021 dataset
Mamba-WheatNet outperformed YOLOv13 in accuracy, recall, and mAP@50
Demonstrated lower computational costs compared to Transformer-based methods
Achieved state-of-the-art performance on the VisDrone2019 dataset

Abstract

Continuous and online estimation of wheat head count is a key factor in guiding wheat production. However, high-precision estimation of wheat head count per unit area still faces challenges due to variety differences and complex environments. To overcome the impact of these unfavorable factors on field yield estimation, the Mamba-WheatNet model is proposed. This model combines the selective state space model (Mamba) and the convolutional neural network (CNN), which has stronger target background distinction and adhesion separation capabilities. At its core, Mamba-WheatNet employs the Bidirectional Selective State Space Module (BSSMBlock), which performs Mamba operations along two spatial dimensions while selectively attending to informative channel features. This design reduces background interference and emphasizes salient wheat head characteristics. To further strengthen feature extraction, we introduce the Residual Depthwise Separable Block (RDSBlock), which leverages lightweight convolutions and multi-branch transformations. The fusion of BSSMBlock and RDSBlock yields the Bidirectional Visual Space Scanning Block (BVSSMBlock), which enhances global context modeling through bidirectional spatial integration. Finally, we conduct comprehensive ablation studies and comparative experiments using the publicly available GWHD-2021 dataset collected via unmanned aerial vehicles. The results show that Mamba-WheatNet outperforms the current state-of-the-art YOLOv13 in terms of accuracy, recall, and mAP@50. Compared to Transformer-based methods (RT-DETR), our method has lower computational cost. To further evaluate the model’s generalization ability, we tested Mamba-WheatNet on the widely used aerial target detection benchmark dataset (VisDrone2019), and the results show that our method also achieves state-of-the-art performance. Therefore, Mamba-WheatNet provides a scalable and accurate solution for wheat ear detection, which helps promote the application of UAVs in smart agriculture.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper