Fusarium head blight (FHB) poses a significant threat to crops, impacting food security and environmental and human health. Traditional FHB management methods are often inefficient and may lead to overuse of chemicals. Utilizing artificial intelligence (AI) for FHB detection enhances disease identification precision, enabling near real-time responses and more calculated interventions.
Early detection is crucial in managing crop diseases. While deep convolutional neural networks have been explored, conventional segmentation frameworks like U-Net face challenges in handling FHB variability. We introduce TransDAE, an innovative approach improving the self-attention mechanism to address spatial and channel dimensions, overcoming the limitations of existing frameworks and refining precision through an inter-scale interaction module.
For training and testing, we collected 12,000 wheat field images using an advanced rotational deep scanning robot. This robot, equipped with a cutting-edge drone featuring a 360-degree rotational camera, provides a unique aerial perspective, enhancing data collection efficiency. The drone's capabilities, including swift image capture from all angles, counterweight system, and calibrated velocity, ensure high-quality images for FHB detection and segmentation. The 360-degree rotational camera adds unprecedented detail to our dataset, advancing our understanding of wheat fields.