<p>Drought stress, as one of the key limiting factors affecting soybean reproductive development, frequently caused photosynthetic product obstruction and seed quality degradation. Deep learning-based crop high-throughput phenotyping techniques combined with image processing demonstrated significant potential in detecting crop biotic and abiotic stresses. Traditional drought recognition methods relying on deep learning models previously exhibited shortcomings such as inadequate three-dimensional feature extraction and fusion capabilities. To address these issues, this study innovatively proposed a drought stress detection method utilizing multi-scale feature fusion for soybean 3D canopy images. First, the ResNet50 network served as the backbone for feature extraction while integrating the SE attention mechanism to enhance feature representation and extract deep semantic features from soybean canopy images. Subsequently, a dedicated multi-scale feature fusion module (MSFM) was designed to effectively aggregate multi-view image features. Finally, the MSAFNet drought stress detection model for soybean 3D canopy images was constructed by mapping aggregated multi-view features to category probability distributions through a fully connected layer and Softmax function. During simulation tests, the model achieved 90.2% recognition accuracy, 91.3% F1-score, and 13.6 fps inference speed. Results demonstrated that the proposed MSAFNet enhanced inference speed while maintaining detection accuracy, providing an effective approach for rapid and precise drought stress identification in soybean cultivation.</p>