Detecting rice diseases using improved lightweight YOLOv8n

Rice is one of the most important staple crops in China, with an annual planting area of approximately 30 million hectares. However, the rice diseases have significantly impacted agricultural production, particularly in the regions with highly intensive farming and a high cropping index. The increasing prevalence of rice diseases has threatened the yield and food security. Early and accurate detection of rice diseases is often required for the effective control of rice diseases. However, several challenges still remain in the existing detection of embedded edge devices, such as the high computational demands of deep learning. This study aims to detect rice diseases using improved lightweight YOLOv8. Spot features of disease were extracted to enhance the detection accuracy in complex field environments. A diverse dataset of rice disease images was systematically collected from real-world fields. The image dataset also included the three major rice diseases: Rice Blast, Bacterial Blight, and Brown Spot. A strong foundation was provided to train and evaluate the deep learning models. In order to improve detection accuracy and computational efficiency, the lightweight model (YOLOv8-DiDL) was proposed to identify the rice disease using YOLOv8n. The key modifications were introduced to enhance the performance of the improved model. Firstly, an Inverted Residual Mobile Block (iRBM) was integrated into the backbone of convolutional modules. The micro disease features were then captured to promote the precision of detection on the small lesions. Secondly, a Deformable Convolutional Network (DCNv2) was incorporated to optimize the geometric size of disease symptoms. Stable performance of detection was achieved in complex and dynamic environments. Thirdly, a Dynamic Sample (DySample) operator was applied to reduce the computational complexity in real-world conditions. The model parameters were minimized for the computational overhead. The improved model was more efficient for the deployment of the resource-limited edge devices. Lastly, the standard Spatial Pyramid Pooling Fast (SPPF) module was replaced with a Large Separable Kernel Attention (LSKA) module. Multiple scale feature was fused in the pooling layer to recognize the diseases over the different scales and lighting conditions. A series of experiments were performed on the standardized platform. The results demonstrate that the improved YOLOv8-DiDL model was achieved with an accuracy of 91.4%, a recall of 83.5%, a mean average precision (mAP) of 90.8%, a parameter count of 2,270,553, and a model weight of only 7.5 MB. Compared with the baseline YOLOv8n network, the improvements were 7.0% in accuracy, 0.5% in recall, and 2.5% in mAP, while simultaneously the model weights were reduced by 9.7% and floating-point operations per second by 7.4%. A comparison was also made on the backbone network, heatmaps, and full-process feature maps. The high effectiveness was found after modifications. Each enhancement positively contributed to both detection accuracy and computational efficiency, thus enhancing the real-time detection of small disease spots. The improved model was achieved with high accuracy and lower computational costs, thus making it feasible for real-world agricultural applications. The reliability of the improved model was validated to detect the rice diseases. An advanced approach was then provided for the precision management in rice fields. Both detection accuracy and deployment efficiency were improved for real-time disease monitoring in intelligent agriculture. The finding can also be further extended into the detection of crop diseases in smart and sustainable agriculture.