The origin of seeds is a crucial environmental factor that significantly impacts crop production. Accurate identification of seed origin holds immense importance for ensuring traceability in the seed industry. Currently, traditional methods used for identifying the origin of maize seeds involve mineral element analysis and isotope fingerprinting, which are laborious, destructive, time-consuming, and suffer from various limitations. In this experiment, near-infrared spectroscopy was employed to collect 1360 maize seeds belonging to 12 different varieties from 8 distinct origins. Spectral information within the range of 11,550–3950 cm⁻¹ was analyzed while eliminating multiple interferences through first-order derivative combined with standard normal transform (SNV). The processed one-dimensional spectral data were then transformed into three-dimensional spectral maps using Gram’s Angle Field (GAF) to be used as input values along with the VGG-19 network model. Additionally, a convolution layer with a step size of 1 × 1 and the padding value set at 1 was added, while pooling layers had a step size of 2 × 2. A batch size of 48 and learning rate set at 10⁻⁸ were utilized while incorporating the Dropout mechanism to prevent model overfitting. This resulted in the construction of the GAF-VGG network model which successfully decoded the output into accurate place-of-origin labels for maize seed detection. The findings suggest that the GAF-VGG network model exhibits significantly superior performance compared to both the original data and the PCA-based origin identification model in terms of accuracy, recall, specificity, and precision (96.81%, 97.23%, 95.35%, and 95.12%, respectively). The GAF-VGGNet model effectively captures the NIR features of different origins of maize seeds without requiring feature wavelength extraction, thereby reducing training time and enhancing accuracy in identifying maize seed origin. Moreover, it simplifies near-infrared (NIR) spectral modeling complexity and presents a novel approach to maize seed origin identification and traceability analysis.

The cover of corn stover has a significant effect on the emergence and growth of soybean seedlings. Detecting corn stover covers is crucial for assessing the extent of no-till farming and determining subsidies for stover return; however, challenges such as complex backgrounds, lighting conditions, and camera angles hinder the detection of corn stover coverage. To address these issues, this study focuses on corn stover and proposes an innovative method with which to extract corn stalks in the field, operating an unmanned aerial vehicle (UAV) platform and a U-Net model. This method combines semantic segmentation principles with image detection techniques to form an encoder–decoder network structure. The model utilizes transfer learning by replacing the encoder with the first five layers of the VGG19 network to extract essential features from stalk images. Additionally, it incorporates a concurrent bilinear attention module (CBAM) convolutional attention mechanism to improve segmentation performance for intricate edges of broken stalks. A U-Net-based semantic segmentation model was constructed specifically for extracting field corn stalks. The study also explores how different data sizes affect stalk segmentation results. Experimental results prove that our algorithm achieves 93.87% accuracy in segmenting and extracting corn stalks from images with complex backgrounds, outperforming U-Net, SegNet, and ResNet models. These findings indicate that our new algorithm effectively segments corn stalks in fields with intricate backgrounds, providing a technical reference for detecting stalk cover in not only corn but also other crops.

Accurate identification of the second and third compound leaf periods of soybean seedlings is a prerequisite to ensure that soybeans are chemically weeded after seedling at the optimal application period. Accurate identification of the soybean seedling period is susceptible to natural light and complex field background factors. A transfer learning-based Swin-T (Swin Transformer) network is proposed to recognize different stages of the soybean seedling stage. A drone was used to collect images of soybeans at the true leaf stage, the first compound leaf stage, the second compound leaf stage, and the third compound leaf stage, and data enhancement methods such as image rotation and brightness enhancement were used to expand the dataset, simulate the drone’s collection of images at different shooting angles and weather conditions, and enhance the adaptability of the model. The field environment and shooting equipment directly affect the quality of the captured images, and in order to test the anti-interference ability of different models, the Gaussian blur method was used to blur the images of the test set to different degrees. The Swin-T model was optimized by introducing transfer learning and combining hyperparameter combination experiments and optimizer selection experiments. The performance of the optimized Swin-T model was compared with the MobileNetV2, ResNet50, AlexNet, GoogleNet, and VGG16Net models. The results show that the optimized Swin-T model has an average accuracy of 98.38% in the test set, which is an improvement of 11.25%, 12.62%, 10.75%, 1.00%, and 0.63% compared with the MobileNetV2, ResNet50, AlexNet, GoogleNet, and VGG16Net models, respectively. The optimized Swin-T model is best in terms of recall and F1 score. In the performance degradation test of the motion blur level model, the maximum degradation accuracy, overall degradation index, and average degradation index of the optimized Swin-T model were 87.77%, 6.54%, and 2.18%, respectively. The maximum degradation accuracy was 7.02%, 7.48%, 10.15%, 3.56%, and 2.5% higher than the MobileNetV2, ResNet50, AlexNet, GoogleNet, and VGG16Net models, respectively. In the performance degradation test of the Gaussian fuzzy level models, the maximum degradation accuracy, overall degradation index, and average degradation index of the optimized Swin-T model were 94.3%, 3.85%, and 1.285%, respectively. Compared with the MobileNetV2, ResNet50, AlexNet, GoogleNet, and VGG16Net models, the maximum degradation accuracy was 12.13%, 15.98%, 16.7%, 2.2%, and 1.5% higher, respectively. Taking into account various degradation indicators, the Swin-T model can still maintain high recognition accuracy and demonstrate good anti-interference ability even when inputting blurry images caused by interference in shooting. It can meet the recognition of different growth stages of soybean seedlings in complex environments, providing a basis for post-seedling chemical weed control during the second and third compound leaf stages of soybeans.