.png)
.svg)
.svg)
Rice is one of the three main staple foods in China. The field management of rice paddies is complex and highly repetitive (such as spraying pesticides and weeding) and extremely labor-intensive, imposing a significant burden on workers. Suzhou Botian uses Baidu's deep learning platform, PaddlePaddle, to enable tractors and agricultural robots to learn visual navigation. They can adjust their course in real-time based on the planting conditions of rice seedlings, avoiding crushing the seedlings and thus better maintaining and managing them. This improvement has made the dream of "saving labor while significantly increasing crop yields" a reality.
Suzhou Botian Automation Technology Co., Ltd. is a high-tech enterprise specializing in the research and development and industrialization of intelligent agricultural equipment technology. Adhering to the development concept of "Technological Innovation, Intelligent Leadership, Crowdsourced Sharing, and Machines Assisting Humans," the company has creatively developed intelligent agricultural machinery equipment such as plant protection robots, fruit and vegetable picking robots, weeding robots, facility agriculture robots, intelligent agricultural machinery management systems, and agricultural machinery auxiliary navigation systems. Many of its products fill domestic gaps and are widely used in more than 20 provinces and cities across the country.
Rice is planted in rows, with rows being approximately parallel to each other. Therefore, the basis for realizing visual automatic navigation of agricultural machinery lies in accurately detecting the centerline of the seedling rows in real-time. Although traditional image processing algorithms can also extract the centerline of seedling rows, the unstructured environment of paddy fields under natural lighting, the differences in image brightness at different times and under different weather conditions, the presence of floating plants and algae with similar characteristics to seedlings, sporadic missing seedlings, and reflections pose significant challenges to the robustness of traditional algorithms, making it difficult to ensure accuracy.
To solve this problem, Suzhou Botian's technical staff conducted a comprehensive analysis of the characteristics of paddy field images and developed an automatic detection system for paddy field navigation lines based on Baidu's deep learning platform, PaddlePaddle. They used the ICNet model from PaddleSeg, a PaddlePaddle image segmentation development kit, to segment the seedlings by row from the background, and based on this, accurately extract the centerline of the seedling rows.
Technicians collected 400 images each of rice seedlings one week and three weeks after transplanting at different times of the day, and varied the exposure time at each time to collect images of different brightness.
The agricultural machinery visual navigation task is somewhat similar to autonomous driving, requiring high real-time performance while ensuring a certain level of accuracy and being deployable on mobile devices such as embedded systems. The ICNet, one of the mainstream semantic segmentation networks supported by PaddleSeg, is a lightweight semantic segmentation network designed for low memory and high real-time application scenarios like autonomous driving, making it very suitable for agricultural machinery visual navigation. Compared to high-accuracy networks like the U-Net series and DeepLab series, ICNet significantly reduces prediction time and memory usage while maintaining a minor reduction in accuracy, achieving real-time results on high-resolution images like 1024*2048 pixels.
During model training, hyperparameters such as the optimizer, number of iterations, and image cropping size can be customized by modifying the relevant configuration files. Additionally, PaddlePaddle includes an image enhancement module, which can be adjusted through configuration files to decide whether to use image enhancement and modify image enhancement strategies, allowing for the training of a network with strong generalization capabilities even with insufficient data. The model input size is set to 640 x 512, and once configured, the training script is invoked to train the model.
During network training, parameters are saved every five epochs. The training loss values and validation accuracy are viewed using PaddlePaddle's visualization tool VisualDL, and the model with the highest validation accuracy is selected to prevent overfitting.
VisualDL visualization graph:
Phase Five: Testing Results
Using PaddleSeg, Suzhou Botian's agricultural robots can now accurately segment the seedlings from the background despite interference, extract the outer contours and feature points of the original image, and subsequently extract the centerline of the middle 4-5 rows of seedlings, laying a solid foundation for agricultural machinery visual navigation. Paired with GPS, Suzhou Botian's agricultural robots have achieved fully automated navigation from the warehouse to the field, significantly reducing labor and resource investment, and ensuring the efficiency and health of farmers' work.
Segmentation results from left to right: segmentation result diagram, original image feature point extraction diagram, seedling row centerline extraction result diagram.
Using the PaddlePaddle platform, Suzhou Botian's technicians achieved visual navigation line extraction in complex paddy field environments with an accuracy rate of over 95%. The time taken to process each image frame (including ICNet network segmentation prediction time and subsequent navigation line extraction time) is about 300ms, meeting the speed requirements of agricultural machinery operations and freeing farmers from repetitive and tedious labor. The most significant improvement is the accuracy of navigation line extraction. Before using PaddleSeg, Suzhou Botian tried various other methods, but accuracy could not be guaranteed under special conditions. Using PaddleSeg-based navigation line extraction, the average angle error in the test images was reduced from 3.27° to 1.65°, with angle errors greater than 5° being rare even under conditions such as reflection or missing plants.
The developers at Botian praised, "PaddleSeg's semantic segmentation library is easy to use, deploy, and feature-rich. It saves a lot of time wasted on 'reinventing the wheel' by allowing the selection of different segmentation models and configuration parameters based on actual needs and test results. This not only simplified our work but also truly realized a technological innovation, accelerating the pace of agricultural machinery intelligence development."
Suzhou Botian's application provides accurate, efficient, and reliable technical support for autonomous navigation agricultural robots, making significant strides towards the great vision of intelligent agricultural machinery and professional farmers, and promoting the development of precision agriculture in China. Next, Suzhou Botian plans to apply PaddlePaddle in greenhouse environments for tasks such as fruit and vegetable picking and intelligent inspection, hoping that more agricultural practitioners can genuinely experience the convenience brought by smart agriculture in the future.
