In the heart of China’s Hebei Agricultural University, a groundbreaking method is being developed to revolutionize the way we count wheat seedlings, a task that has long been a labor-intensive and time-consuming process. Menghan Li, a leading researcher from the College of Mechanical and Electrical Engineering, has spearheaded a project that leverages the power of deep learning and laser labeling to automate this crucial agricultural task. The research, published in the journal *Smart Agricultural Technology* (translated as *智能农业技术*), promises to bring significant efficiencies to the agricultural sector, with potential ripple effects across the global food supply chain.
The traditional methods of counting wheat seedlings are not only tedious but also prone to human error. “Manual counting is subjective and inefficient,” Li explains. “It’s a process that can take hours, even days, and the results can vary greatly depending on the individual doing the counting.” To address these challenges, Li and her team have developed a two-stage convolutional neural network (CNN) that automates the counting process with remarkable accuracy.
The first stage of the process involves labeling wheat seedlings with a laser and capturing the image data using a self-developed data acquisition device. The team then employs the SCNet model for wheat seedling segmentation, using ResNet50 as the backbone network and incorporating the Convolutional Block Attention Module (CBAM) for feature enhancement. “This step is crucial because it allows us to focus on the wheat seedlings and ignore the background noise,” Li notes.
In the second stage, the team uses the SS-DeepLabV3+ model for laser point segmentation. Here, MobileNetV2 serves as the encoder, with the Squeeze-and-Excitation (SE) attention mechanism and self-attention mechanism enhancing the model’s feature representation. The final step involves counting the number of laser points, which corresponds to the number of wheat seedlings.
The results of this two-stage algorithm are impressive. Evaluation metrics such as mean Pixel Accuracy (mPA), Recall, and mean Intersection to Union (mIoU) all show significant improvements over existing methods. This new approach not only saves time and labor but also provides a more accurate count, which is essential for subsequent yield prediction and overall farm management.
The implications of this research extend far beyond the fields of Hebei. In an era where precision agriculture is becoming increasingly important, tools like Li’s automated counting system can help farmers optimize their resources, reduce waste, and increase productivity. “This technology has the potential to transform the way we approach agriculture,” Li says. “By providing more accurate and timely data, we can help farmers make better decisions and ultimately improve the efficiency of the entire food supply chain.”
As the world grapples with the challenges of feeding a growing population, innovations like this one are more important than ever. The research published in *Smart Agricultural Technology* offers a glimpse into the future of agriculture, where technology and tradition intersect to create more sustainable and efficient farming practices. With further development and widespread adoption, this two-stage convolutional neural network could become a cornerstone of modern agriculture, shaping the way we grow and harvest our food for years to come.