Abstract: Abstract: Rice panicle blast is one of the most serious diseases in the period of rice growth. To effectively identify the rice panicle blast is one of the important prerequisites for rice disease controlling. In this study, a novel identification method for panicle blast based on hyperspectral imaging technology is proposed. The method applies a deep convolutional neural network model GoogLeNet to learn the representation of hyperspectral image data and the binary panicle blast/non-blast classifier is trained as well in a unified framework. The GoogLeNet is 22-layer deep convolutional neural network, which repeatedly stacks basic Inception module to deepen and widen the network to enhance its representation power. The core Inception architecture uses a series of kernel filters of different sizes in order to handle multiple scales macro structure and all of filter parameters are learned. In our GoogLeNet model for the panicle blast identification, the filter sizes are set to 1×1, 3×3 and 5×5 based on the consideration of lesion microstructure size rendered on the rice spike. In order to reduce the expensive computing cost of 3×3 and 5×5 convolutions, an extra 1×1 convolution is used to reduce the map dimension in each branch of Inception module before 3×3 and 5×5 convolutions. Further, all the output filter banks are concatenated into a single output vector forming the input of the next stage. As these Inception modules are stacked on top of each other, features of higher abstraction are captured by higher layers. Finally, an average pooling layer plus a fully connected layer is stacked on the last Inception module and a softmax based classifier is used to predict the panicle blast. From the statement, feature and classifier learning are seamlessly integrated in a unified framework and both of them are trained jointly under the supervision of blast label, which makes the two reach the harmoniously optimal state and helps to improve the blast prediction performance. To verify the acclaim of the proposed GoogLeNet method, a total of 1 467 fresh rice panicles covering more than 71 cultivars are collected from an experimental field for the performance evaluation. The experimental field is located in regional testing area for the evaluation of rice cultivars in Guangdong Province. Therefore, all the rice plants in this area are naturally inoculated as the area is a typical source of rice blast fungus. The hyperspectral images of all the rice panicles are acquired using outdoor portable GaiaField-F-V10 imaging spectrometer. In consideration that the spatial resolution is large, we coarsely crop the background area. Then the average spectrum images are computed, acting as the original input of the deep GoogleNet network. Two-class label of hyperspectral image sample is determined by plant protection expert according to the description of blast infection. In our experiments, totally 200 samples are randomly selected for test, with 100 for infected and non-infected class respectively. The rest are for training. When the training samples are scarce, deep GoogLeNet model is easily trapped in the overfitting, worsening the panicle blast prediction performance. To this end, we proposed 2 data augmentation methods, i.e., the method of randomly abandoning single band and the method of randomly translating luminance of average hyperspectral image. The combination of 2 methods can produce hundreds of thousands of data sample pairs. The rich and diverse samples are used to train the deep convolutional model to reduce the overfitting and improve the prediction results. Experimental results show that the proposed GoogLeNet based method achieves a high classification accuracy of 92.0%. This result is much better than the recent state-of-art BoSW (bag of spectra words) method, demonstrating the proposed GoogLeNet method together with the 2 data augmentation techniques solves the panicle blast identification problem under the situation of outdoor hyperspectral image collection. Moreover, the proposed GoogleNet BoSW based method demonstrates strong robustness to rice cultivars, which is vital for the wide and practical application. This research improves the classification accuracy of rice panicle blast identification and overcomes the difficulty caused by the hyperspectral image collection under the natural light outdoor. This work will advance the research of panicle blast identification to the practical application of production with a big step.