Abstract: There are relatively few studies on the spray quality of citrus tree canopy in China. In most cases, the method is that farmers observe the spray quality of citrus tree canopy up close in the orchard, which wastes manpower and material resources. Moreover, the observation effect is not ideal and may cause harm to human safety. Domestic conditions allow some orchards to judge the spray effect by observing the water-sensitive paper. This method does save a lot of manpower, and efficiency is significantly improved compared with the previous ones. However, this method is affected by the external environment, and the naked eyes cannot correctly determine the spray quality, which will cause certain errors. Therefore, this study explored new ways to solve these problems. In recent years, thermal imaging technology had shown great research promise in some emerging research areas, especially in agricultural production. In the field of precision agriculture, for example, high-resolution thermal imaging cameras, with the aid of advanced aerial photography technology, can quickly capture thermal images of the canopy before and after spray, which provides new ideas for precision agriculture spray detection technology. This study combined thermal imaging technology and computer vision technology to identify and classify the spray conditions of the plant canopy and accurately detected the spray quality of the leaves, which avoided the waste of pesticides and reduced the manual re-examination steps. Thereby reducing agricultural production costs and improving the economic benefits of agricultural products. In this study, the canopy of a citrus tree was used as a thermal image acquisition area. Inevitably, there are problems such as high noise, low contrast, and blurred feature information in the thermal image acquisition process. In response to the above problems, this study preprocessed the acquired thermal images and eliminated unreasonable data, and set the original data set to two labels (sprayed and unsprayed), and divided them into the training set and the test set. An improved SPP-Net-Inception-v4 model based on the Inception-v4 model and the SPP-Net target detection algorithm was proposed to achieve multi-feature fusion to enhance the feature extraction effect. The model took the construction of a sparse network structure to generate dense data as the core design idea. By introducing the Inception and Reduction modules, the feature description bottleneck problem was reduced; further, SPP-Net (Spatial Pyramid Pooling Network) was innovatively connected between the convolutional layer and the fully connected layer pooling network), which aimed to extract fixed-length feature vectors through the pyramid space pooling method to achieve the fusion of multi-scale features and extraction enhancement effect. Compared with the two models of Inception-v4 and ResNet-152, the experimental results showed that the accuracy of the improved SPP-Net-Inception-v4 model test set was 95.07%, which was 1.58% higher than the accuracy of the original Inception-v4 model and 3.26% higher than that of ResNet-152. Compared with Inception-v4 and ResNet-152, the SPP-Net-Inception-v4 model reduced the model size by 13% and 24%, respectively. The SPP-Net-Inception-v4 model could be used to detect the spray quality of chemicals in citrus fruit trees quickly as well as could improve the economic benefits of agricultural production, and would provide a reference for the further improvement of pesticide detection technology in precision agriculture.