RPH-Counter: Field detection and counting of rice planthoppers using a fully convolutional network with object-level supervision

Rice planthoppers are among the most severe migratory pests affecting rice, characterized by small size and rapid reproduction, leading to sudden and explosive outbreaks. Therefore, timely and accurate monitoring of rice planthopper populations is crucial. Applying machine vision to field monitoring of rice planthoppers can reduce labor and material costs. Existing literature lacks research on field detection and counting of rice planthoppers, and general detection and counting methods suffer from performance degradation in complex environments. In this study, we propose the Rice Planthopper Counter (RPH-Counter), a novel detection and counting architecture. The model is a simple Fully Convolutional Network (FCN). Initially, we propose the Object Counting loss (OC loss), which includes four sub-loss functions that compel the FCN to learn each object's center and boundary positions while constraining false positives. After training, the FCN can predict a separate spot for each rice planthopper, achieving precise localization and counting of the pests. Then, we propose the Self-Attention Feature Pyramid Network (SAFPN) by adding additional Spatial Self-Attention (SSA) modules at the lateral connections of C3 to C5, enhancing the model's performance in complex environments at a lower computational cost. We collected a large-scale field rice planthopper dataset, containing approximately 140,000 annotated rice planthoppers. The evaluation metrics are localization accuracy and counting error. Experimental results show that the RPH-Counter, with lower computational complexity, significantly improves performance, achieving an F1 score of 92.36%, a Mean Absolute Error (MAE) of only 2.40, and an R-squared (R2) of 0.985. Compared to the state-of-the-art object detectors, the F1 score improved by 8.62%, and the counting error decreased by 61%. Compared to the state-of-the-art density estimation methods, the counting error decreased by 23%, with precise localization ability and multi-class expandability. This method offers a new research approach and promising direction for field pest counting and pest population monitoring.