Wheat seedling counting is one of the most crucial field surveys at the emergence stage. However, the challenges remain, such as small size, dense distribution, and mutual occlusion of wheat seedlings. This study aims to accurately detect and count the wheat seedlings in complex field environments. An improved algorithm (named DSM-YOLOv10) was proposed using YOLOv10n. Firstly, global annotation was adopted rather than only local features during data labeling. The overall features of wheat seedlings were learned and extracted to effectively avoid the feature loss caused by overlapping and occlusion. Secondly, a DCC2f module was constructed to replace the C2f module with the dual convolution (DualConv) in the backbone network. Grouped and pointwise convolutions were concurrently deployed to combine with the residual connections. The gradient vanishing was alleviated in deep-layer training. A more complete transmission of cross-layer features was realized to minimize the redundant convolution. Computational efficiency was improved to facilitate subsequent real-time deployment on mobile edge devices. More efficient features were provided for the subsequent processing. Furthermore, a semantics and detail infusion (SDI) module was introduced to explicitly inject the high-level semantic information into low-level detail features using cross-layer attention and Hadamard products. Detail enhancement was obtained under semantic guidance to avoid simple feature stacking without deep information interaction in the conventional concatenation modules. Consequently, the performance was improved to decouple and reuse features in the overlapping regions of wheat seedlings. Finally, a multi-scale dilated attention (MSDA) mechanism was adopted to incorporate the multiple dilation rates. Multi-scale semantic information was effectively aggregated to integrate the local details with contextual information. Furthermore, the intersecting and overlapping wheat seedlings were enhanced after extraction. Moreover, the complex redundancy or additional computational cost was efficiently reduced in the self-attention mechanism. Experimental results demonstrate that the DSM-YOLOv10 model achieved precise detection of wheat seedlings under multiple complex scenarios. The mean average precision (mAP), precision (P), recall (R), and F1 score were 91.4%, 85.2%, 81.7%, and 83.4%, respectively, which was improved by 5.0, 2.6, 5.4, and 4.1 percentage points, respectively, compared with the original YOLOv10n model. Furthermore, the parameter count and floating point operations (FLOPs) were reduced by 4.7% and 10.7%, respectively. The real-time detection was also realized with an inference time of 13.2 ms (approximately 76 frames per second). Compared with the RetinaNet, Faster-RCNN, SSD, and YOLOv8n models, the DSM-YOLOv10 model exhibited the best performance in detection. The mAP values were 35.8, 24.5, 30.7, and 8.6 percentage points higher, respectively; whereas, the number of parameters was reduced by 87.4%, 96.2%, 90.0%, and 15.2%, respectively; and the FLOPs were decreased by 86.3%, 96.0%, 88.1%, and 7.4%, respectively. The DSM-YOLOv10 model also demonstrated excellent recognition of the small objects, with a mAP of 86.3%, particularly in overlapping seedling scenarios. Both its missed detection rate and false detection rate remained at the low levels, at 9.1% and 3.8%, respectively. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were 0.92, 5.68 plants, and 4.33 plants, respectively, in the seedling counting task using the DSM-YOLOv10 model. Compared with the YOLOv10n, the R² increased by 6.98%, while the RMSE and MAE decreased by 35.23% and 40.44%, respectively, indicating the higher counting accuracy and robustness. This finding can effectively improve the detection and counting performance of wheat seedlings in complex scenarios. The improved model was also featured by fewer parameters and lower FLOPs. The finding can provide support for field data acquisition in smart agriculture.

High accuracy and efficiency can often be required to assess the freeze damage of the winter wheat. However, the traditional field phenotypic identification can be confined to strong subjectivity, serious deviations between laboratory simulations and field data, and the lengthy cycle of physiological and biochemical detection. In this study, a freeze damage classification was proposed to fuse the vegetation indices, texture, and color features. A freeze damage model was also constructed to combine with the machine learning. The unmanned aerial vehicles (UAVs) were utilized to collect the multispectral and visible light imagery. Specifically, the UAVs were equipped with multispectral sensors and high-resolution visible light cameras. The optimal weather (clear skies, stable and light) was employed to ensure the data quality. A total of 15 vegetation indices were extracted, including the normalized difference vegetation index (NDVI), ratio vegetation index (RVI), excess blue index (ExB), and soil adjusted vegetation index (SAVI). The spectral reflectance was also selected from the specific wavelength bands in order to effectively represent the vegetation’s photosynthetic capacity and canopy health status. Furthermore, the gray-level co-occurrence matrix (GLCM) was employed in texture feature analysis. Some parameters were calculated to quantify the spatial distribution of the pixel gray levels in the imagery, such as the mean, variance, contrast, and homogeneity from the GLCM. These texture features were then captured for the subtle structural variations in the wheat canopy caused by freeze damage. Additionally, the color features were derived from the RGB color space of the visible light images. Statistical parameters were computed to represent the visual color of the winter wheat under freeze damage, such as the color mean, standard deviation, and color moment. Feature screening was conducted using the Pearson correlation coefficient (PCC) method and the minimum redundancy feature selection algorithm (mRMR). Firstly, the Pearson correlation analysis was used to measure the linear correlation between each feature and freeze damage levels, thus eliminating the features with a low correlation. Subsequently, the minimum redundancy feature selection was optimized the feature set, in order to balance the feature relevance and redundancy. The final features were selected for the informative and non-repetitive ones. Four classification models were then constructed: random forest (RF), support vector machine (SVM), backpropagation neural network (BPNN), and eXtreme gradient boosting (XGBoost). Each model was also trained with tunable parameters. For example, the XGBoost model was utilized to adjust the parameters, such as the learning rate, the number of estimators, and the maximum depth after cross-validation. The experimental results demonstrated that the multi-source feature fusion significantly improved the performance of the model. Among them, eXtreme Gradient Boosting (XGBoost) showed the optimal performance under the fusion of vegetation indices, texture, and color features. The accuracy, recall, precision, and F1-score reached 0.73, 0.68, 0.71, and 0.69, respectively. Further analysis revealed that the texture features shared the highest correlation with the manually surveyed levels of the freeze damage. The structural damages of the canopy were also sensitive to capture, as the freeze damage often disrupted the orderly arrangement of the wheat leaves and stems. The texture features were effectively detected after capture. The multi-source remote sensing data fusion was validated to assess the freeze damage. The spectral, texture, and color information were integrated to precisely monitor the freeze damage of the winter wheat for disaster prevention decision-making. Remote sensing and machine learning were also applied to assess agricultural disasters in smart agriculture. The finding can serve as a practical reference for crop stress monitoring in the disaster early warning of intelligent agriculture.