Detecting potholes on unstructured farm roads is a critical prerequisite for the autonomous navigation of agricultural machinery. Yet the challenge remains, due to complex field environments, blurred target boundaries, irregular morphologies, and significant scale variations. In this study, a pothole detection model, named POT-YOLOv11n, was proposed using the baseline YOLOv11n architecture. Three components were incorporated: A Multi-scale Edge Information Enhancement module was integrated multi-branch convolutional layers with channel-wise attention mechanisms to strengthen edge feature extraction and boundary discrimination for potholes with vague contours; A Pyramid Pooling and Large Kernel Attention Fusion module was combined large kernel separable attention with spatial pyramid pooling to enhance multi-scale contextual feature capture; And a Feature Screening and Contextual Anchor Attention mechanism was introduced in the neck network to optimize feature fusion and contextual modeling. Salient features were focused on to suppress the irrelevant information from complex backgrounds. In static evaluation, a pothole dataset of a farm road was constructed using diverse samples with varying scales, irregular shapes, and different lighting. The improved POT-YOLOv11n model achieved a mean Average Precision of 83.20%, indicating an improvement of 2.47 percentage points over the baseline YOLOv11n model. An F1-score of 80.30% was obtained to balance the precision and recall performance. Compact architecture was maintained with only 2.22 million parameters, with an inference speed of 248.83 frames per second, thereby meeting real-time requirements for autonomous navigation. In dynamic evaluation, field tests were conducted on a representative farm road segment with typical potholes. Practical applicability was assessed under realistic operation, where the agricultural vehicle was operated at three conventional working speeds, including low speed at 5 km/h, medium speed at 10 km/h, and high speed at 15 km/h. The better performance was achieved in the continuous image acquisition and online detection. As the vehicle speed increased, motion blur effects intensified, leading to a decline in detection accuracy. At a low speed of 5 km/h, the mean Average Precision of 80.8% was slightly lower than the static test with precision and recall of 82.1% and 79.6%, respectively. While the F1-score remained consistent with static testing, indicating effective transferability of detection during real-world deployment; At medium speed of 10 km/h, the mean Average Precision decreased to 79.8% with the precision and recall of 80.9% and 78.8%, respectively; At high speed of 15 km/h, the mean Average Precision further declined to 75.5% with the recall decreasing to 74.2%, indicating increased missed detections under severe motion blur, yet a mean Average Precision of 75.5% was obtained under such high-speed conditions. Therefore, the multi-scale edge enhancement module also provided for the compensation against motion blurs, preserving detection performance beyond what conventional architectures. The POT-YOLOv11n model achieved effective performance in detecting potholes on unstructured farm roads over different operating speeds. Boundary ambiguity, scale variation, and motion blur were solved using multi-scale edge enhancement, large kernel attention fusion, and feature screening mechanisms. While a lightweight architecture and high inference speed were suitable for practical deployment, the consistent performance in the static and dynamic tests, particularly the robust data under low and medium speeds, and the compensated performance at high speed for the practical reliability of the approach. Therefore, the POT-YOLOv11n model can provide reliable support for the autonomous navigation of agricultural machinery in complex field environments, thus contributing to smart agriculture and farming. A combination of accuracy, efficiency, and robustness was obtained to position motion blur. The finding can provide a promising solution for real-time obstacle detection in vision navigation systems in smart agriculture.

Quantitative evaluation standards are often required to accurately predict sprouting regeneration, particularly for cutting quality. However, the coppicing surface targets cannot be recognized in complex fields during Caragana korshinskii shrub coppicing in the arid regions of Northwest China. In this study, a quantitative evaluation was proposed for sprouting regeneration, according to the synergistic association between cutting morphologies and agronomic traits. Accurate recognition of coppicing surfaces was achieved in unstructured field environments. A segmentation network was constructed, named WoodGrainNet. Firstly, the global context modelling and target localization were enhanced via a semantic stream using a lightweight Transformer architecture. Secondly, a frequency-domain stream was incorporated with the Haar discrete wavelet transform to effectively suppress abiotic background interference for the texture representation of the targets. Simultaneously, a shape stream was combined with a differentiable Sobel operator. Explicit constraints were also applied to high-frequency gradient regions. Thereby, the adjacent cut boundaries were delineated to effectively alleviate the adhesion of dense targets. Three-stream features were synergistically simulated for the deep integration. The WoodGrainNet significantly improved the segmentation robustness and instance discrimination of the improved model in complex scenes, particularly at the level of underlying physical feature representation. Accurate segmentation was achieved to automatically extract a key geometric phenotypic indicator for coppicing quality—coppicing surface circularity (C). Accordingly, the coppicing quality grading and sprouting prediction were established after evaluation. The experimental results demonstrate that the better performance of WoodGrainNet was achieved to balance high accuracy and real-time processing, with a mean Intersection over Union (mIoU) of 86.99% and an inference speed of 51.78 frames/s. The performance was significantly superior to mainstream networks, such as DeepLabV3+, effectively recognizing tiny coppicing targets. In terms of agronomic analysis, statistical results revealed that there was a significant positive correlation between the morphological quality of the coppicing surface and the sprouting potential of lateral branches. Notably, the coefficient of determination (R²) between the coppicing surface circularity (C) and the number of lateral branch sprouts reached 0.764 at the 5th week (N). The circularity served as an important morphological indicator to evaluate coppicing quality and sprouting regeneration. Field tests were conducted to verify its feasibility for engineering applications. The improved model was deployed on a mobile intelligent system. A discrimination accuracy of 90.8% was achieved for coppicing quality grades under mobile working conditions. Sprouting prediction trend was highly consistent with the measured ones, indicating its potential to replace manual inspection and subjective empirical judgment. The pixel-level phenotypic analysis of the coppicing surface was effectively transformed to predict the single-plant regeneration. The finding can provide a theoretical basis and technical support for the quality evaluation, parameter monitoring, and precise detection of ecological shrub forests in arid regions.