林业业务图像的识别分类有利于林业管理部门对相关事件作出合理的处置方案及指挥调度决策，从而充分发挥护林员的作用，提升森林管护的水平，达到保护森林资源和生态安全的目的。

提出了一种针对林业业务图像基于迁移学习的卷积神经网络（convolutional neural networks）自动分类模型。在经过大规模辅助图像数据集ImageNet预训练的4种卷积神经网络模型的基础上，使用林业业务图像数据对训练好的模型进行迁移学习，采用新的全连接层取代原始的全连接层，其他层参数保持不变。

在建立的4个类别林业业务图像数据集上，4个预训练卷积神经网络结构的迁移学习模型都具有较高的分类正确率。其中，基于Inception?v3的迁移学习模型识别精度最高，达到96.4%。

利用基于Inception?v3的迁移学习模型进行林业业务图像分类是可行的。相比传统的特征提取识别方法以及其他预训练模型，Inception?v3模型具有很强的分类能力，可以在森林管护中发挥更广泛的应用。

The recognition and classification of forestry business images are helpful for forestry administrative departments to make reasonable disposal plans, instruct others, and communicate decisions for relevant events. Therefore, full information can be provided to the forest protection personnel, which will improve the level of forest management and protection. Then, the protection of forest resources and ecological security can be achieved.

We proposed an automatic classification model of convolutional neural networks based on the transfer learning for forestry business images. Four convolutional neural network models were pre-trained using a large-scale auxiliary image dataset, called ImageNet. Then, the forestry business image data, with a relatively small number of images, were used to transfer the trained model. The classification accuracy was compared and analyzed to select the model with the best classification effect. In this study, four convolutional neural network models were selected.The forestry business image dataset used for the model training and testing contained four classes, including animal death (AD), forest harvesting (FH), forest fire (FF), and forest pests (FP). The dataset included a total of 280 images, with 70 images in each class. The forestry business image dataset was divided into a training and test set, which accounted for 90% and 10%, respectively. The 10-fold cross-validation method was used to obtain the final test accuracy rate. A new fully connected layer was used to replace the original fully connected layer, while the other layers remained unchanged. The transfer learning method was divided into two steps. The Adam optimization algorithm was used to train the models. In addition, to compare the impacts of different data enhancement methods on the classification effect, three experimental scenarios were used in this study.

The transfer learning models that we used with the four pre-trained convolutional neural network structures all had a higher classification accuracy compared with the traditional method. In particular, Inception-v3 had the highest test accuracy rate, reaching 96.4%, and the unit inference time was only 1.38 s. Inception-v1’s accuracy rate was slightly higher than Inception-v2’s. VGG-16 had the lowest accuracy rate, at only 92.1%, and the unit inference time was also the longest, at 2.42 s. Concerning the four categories of forestry business image datasets, the Inception-v3-based transfer learning model had the highest recognition accuracy. Compared with the VGG-16 model, the classification accuracy of Inception-v3 had a better recognition performance at a shorter recognition time. The experimental results concerning the different data enhancement methods showed that, compared with the test accuracy rate without data enhancement, the test accuracy rate after enhancing the data by 4 times was still 96.4% after 6 000 training steps. However, the test accuracy rate was approximately 96.8% between 1 000 and 3 000 steps, which was a slight improvement. Comparing the curves of the models with no-data enhancement and data enhanced 16 times, we found that the test accuracy rate of the enhanced dataset decreased, and as the number of training iterations increased the test accuracy rate became lower. In addition, the fluctuation of the test accuracy rate curve of the dataset enhanced 16 times was more pronounced compared with that of the other two cases. To analyze the “understanding” of the forestry business image dataset based on the Inception-v3 transfer learning model, a confusion matrix was constructed. It was found that FH was more likely to be confused with AD, while FP was more likely to be confused with FH. In contrast, FF was rarely confused with the other classes. There were obvious red flame features on the FF images, which were rare in the images of the other classes, so it was easy to distinguish between them. However, the other three classes did not have such distinguishing features, so there was confusion on some occasions. For the traditional classification methods, the extracted features were manually selected and had poor generalization characteristics. In this study, the traditional methods had a relatively poor classification accuracy concerning forestry business image classification.

Compared with the traditional feature extraction recognition methods, the convolutional neural network model had stronger classification capabilities. By modifying the network structure, the accuracy rate can be effectively improved. Therefore, it is feasible to use convolutional neural network models based on the transfer learning for forestry business image classification. With the improved recognition rate of convolution neural network models, forest rangers can play a greater role in forest protection.