Abstract: Sunflower is one of the major oil crops, valued for its strong tolerance to nutrient-poor soils and drought, making it a preferred crop for the remediation of mildly to moderately saline-alkali soils. Sunflower disk rot, caused by sclerotinia sclerotiorum infection, leads to increased sterile seed rates, reduced yield, and a decline in nutritional components such as kernel protein and oil content, thereby compromising both the edible and commercial value of the seeds. Consequently, efficient and accurate assessment of disk rot severity is crucial for early disease control, precision pesticide application, and yield estimation. To address the inefficiency and subjectivity of traditional disease classification methods, this study proposes an improved YOLOv12n-based model for detecting the severity of sunflower disk rot at the mature stage. First, the C3K2 modules in the backbone and neck networks were replaced with C3K2-RC modules, which integrate receptive-field attention convolution (RFAConv) and coordinate attention (CA) mechanisms to enhance the model's feature extraction capability for different severity grades of disk rot in complex environments. Second, a lightweight upsampling operator, CARFAE, was integrated into the neck network to improve feature reconstruction. Finally, a lightweight shared convolutional detection head with separated batch normalization (LSCSBD) was introduced to improve the detection accuracy and speed for small-scale lesions. Experimental results show that the YOLOv12n-RCL model achieved precision, recall, mAP0.5, and mAP0.5～0.95 of 83.4%, 81.2%, 84.8%, and 50.2%, respectively, representing improvements of 3.8, 3.2, 3.4, and 4.0 percentage points over the baseline model. The number of parameters, computational complexity, and model size were reduced to 2.11 M, 5.7 GFLOPs, and 4.5 MB, corresponding to reductions of 17.9%, 12.3%, and 19.6% compared to the original model. Analysis based on the normalized confusion matrix indicated that the recall for disease severity grades 0, 1, 2, 3, and 4 were 83.0%, 81.2%, 79.1%, 80.7%, and 82.0%, respectively, demonstrating a balanced recognition capability across all five grades. Furthermore, 8.2% of samples with a true label of grade 2 were misclassified as grade 3, and 9.3% of samples with a true label of grade 3 were misclassified as grade 2. Therefore, confusion between severity grades occurred primarily between grades 2 and 3, while for the remaining grades, the model's misclassification rate remained below 10%, reflecting its strong discriminative ability for disk rot at different grades. In tests conducted on 70 images containing 1298 samples of different grades of disk rot, YOLOv12n-RCL achieved only 10 missed detections and 8 false detections. Compared to the baseline YOLOv12n model, it demonstrates superior detection performance and a higher frame rate, reaching 27.5 frames per second (FPS). Visualization and analysis of sunflower disk rot areas based on UAV orthophoto imagery demonstrated that the improved model maintained stable detection performance even in complex scenarios with dense target distribution and leaf occlusion, without significant missed or false detections. In summary, the YOLOv12n-RCL model improves detection accuracy while effectively balancing lightweight design and deployment efficiency. It can serve as an algorithmic reference for the development of intelligent control equipment targeting sunflower disk rot at the mature stage.