Abstract: The traditional manual control measures for apple disease primarily rely on subjective experience, which can lead to arbitrariness and bias. This approach is also prone to pesticide waste and misuse. As apple leaves are one of the high-incidence areas of diseases, achieving natural environment-based identification of apple leaf diseases can provide effective guidance for disease prevention and control. To address the challenge of accurate identification of apple leaf diseases in natural settings, a DenseNet121+EfficientNet model with Focal Loss and Label Smoothing (DEFL) is designed. Given that apple leaf disease lesions are small and exhibit varying spatial characteristics for different diseases, accurately localizing and identifying disease regions is challenging. The DEFL model utilizes two parallel networks: EfficientNet-B0 and DenseNet121 for feature extraction. It combines the semantic and positional information extracted from these networks to enhance the model's fine-grained feature extraction capability. To overcome difficulties in identifying apple leaf samples due to the similarity in features among different diseases, a Focal Loss function, combined with label smoothing strategies, is introduced. This prevents the model from being overly confident about the class of apple leaf samples, focusing on those samples with challenging decision boundaries during training, thereby improving the model's ability to recognize boundary cases and enhancing its robustness and generalization. The DEFL model proposed in this study was tested with 3906 apple leaf disease images in natural scenes. The results show that the overall recognition accuracy of the model proposed in this study is 99.13%, and the mean average accuracy is 98.47%. The results of ablation experiments show that the mean average precision of the model is improved by 7.99% after adding DenseNet121 feature extraction branch to EfficientNet-B0. And, after introducing the focus loss function combined with label smoothing strategy to EfficientNet-B0, the mean average precision of the model is improved by 3.15%. After the combination of the two improvements, the mean average precision of the model increased by 12.29%. Visualization results using Gradient-weighted class activation mapping (Grad-CAM) heatmaps show that the model accurately localizes and focuses on disease regions. The image feature information used by the model for apple leaf disease recognition is reliable. Uniform manifold approximation and projection (UMAP) feature dimension reduction visualization results indicate that the Focal Loss function with label smoothing effectively enhances the model's feature extraction capability, resulting in more distinctive feature information and improved disease recognition. The comparative experimental results show that compared with mainstream recognition models such as ResNet50, Inception V3, and ResNeXt, as well as EfficientNet-B0 models that fuse these three models, the mean average precision of the proposed model has increased by 14.53, 13.17, 14.61, 6.4, 7.71, and 8.91 percentage points, respectively, and the model size has decreased by 18.73, 7.7, 12.2, 83.62, 69.6, and 60.09 MB, respectively. Comparative experiments have shown that the proposed model has the best recognition performance. In practical applications, the DEFL model achieved an overall recognition accuracy of 97.73% and a mean average precision of 95.82% when tested with 1501 apple leaf images collected from apple plantations. In conclusion, the DEFL model proposed in this study enables accurate and rapid identification of apple leaf diseases, providing valuable guidance for apple disease prevention and control.