Field detection of tiny pests from sticky trap images using deep learning in agricultural greenhouse

Abstract Agricultural pest catches on sticky traps can be used for the early detection and identification of hotspots, as well as for estimating relative abundances of adult pests, occurring in greenhouses. This study aimed to construct a detection model for whitefly and thrips from sticky trap images acquired in greenhouse conditions. An end-to-end model, based on the Faster regional-convolutional neural network (R-CNN), termed ‘TPest-RCNN’, was developed to improve the tiny pest detection accuracy. This architecture was trained using a transfer learning strategy on the Common Objects in Context dataset before training on the tiny pest training set to create the TPest-RCNN model. The new model achieved mean F1 score and average precision of 0.944 and 0.952, respectively, on a validation set. The TPest-RCNN model outperformed the Faster R-CNN architecture and other approaches using handcrafted features (color, shape and/or texture) in detecting multiple species from yellow sticky trap images. The test results also showed the model was robust to detect tiny pests on images of different pest densities and light reflections. Using a linear regression between the manual counts and an automatic detection results using the proposed method on images of 41 days, the determination coefficients reached 99.6% and 97.4% for whitefly and thrips, respectively. These results demonstrated that the proposed method could facilitate rapid gathering of information pertaining to numbers of the abundance of tiny pests in greenhouse agriculture and provide a technical reference for pest monitoring and population estimation.