AN IDENTIFICATION METHOD OF PINUS MASSONIANA PEST AREA USING IMPROVED GOOGLENET

The recent years have witnessed the increase of forest biological disasters. The conventional identification methods of Pinus massoniana pest area have the problem of low accuracy. With the help of artificial intelligence and big data technology, this paper proposes an identification method of Pinus massoniana pest area based on improved GoogLeNet. First, five features of Pinus massoniana images are extracted: Color and texture features are extracted respectively using color moments and gray level cooccurrence matrix, and three spectral features are extracted from the relative spectral reflectance of three bands. Then, a network model is built based on improved GoogLeNet. Through transfer learning, the knowledge of GoogLeNet is transferred to the task of identifying the pest area of Pinus massoniana. The improvement lies in the use of multi-scale convolution kernel to extract the distribution characteristics of pests. Finally, activation function and gradient descent algorithm are optimized to improve the performance of pest identification. Experimental dataset, from image sets of Pinus massoniana pest area in Zhejiang Province, China, is used to test the proposed method in TensorFlow framework. The results show that compared with other methods, the proposed network has better performance in identifying Pinus massoniana pest areas. The accuracy and Kappa index are 94.36% and 0.91. Besides, the proposed network has stronger robustness and applicability, which can provide reference for the identification and intelligent diagnosis of plant pests such as Pinus massoniana.