Rapid density estimation of tiny pests from sticky traps using Qpest RCNN in conjunction with UWB-UAV-based IoT framework.

Precision agriculture has long struggled with the surveillance and control of pests. Traditional methods for estimating pest density and distribution through manual reconnaissance are often time-consuming and labor-intensive. To address these challenges, this study proposes a novel farmland detection system utilizing UAVs, yellow sticky traps, and deep learning techniques. The system includes a UAV based on UWB communication and positioning technology to collect picture information of sticky traps arranged in farmland. Moreover, a faster and more accurate Qpest-RCNN model is used to count the number of insects in the collected sticky traps. The bivariate kernel density estimation establishes the pest density distribution map. Regarding the dynamic monitoring of pest density in agricultural fields, it primarily involves four components: reaching the designated area, flight learning, image acquisition, and visual counting of insects and calculation of insect density. Experimental results demonstrate that UAVs require less time to adjust flight posture during image acquisition after undergoing flight learning, resulting in more concise flight trajectories. The Qpest-RCNN model introduces a variety of mechanisms based on the characteristics of the collected sticky trap data set to improve the faster-R-CNN model. We used two data sets separately to train the model, which were collected by sticky traps placed in greenhouses and open-air experimental fields. The data set from the greenhouse is an open-source dataset provided by M. Deserno et. al, map, precision, and recall of model are 0.923, 0.989, and 0.919, respectively. When the data set collected in the experimental farmland is used to train the model, map, precision, and recall of model are 0.781, 0.851, and 0.789, respectively. In the meantime, we explored the effects of species interference on visual insect statistics and the optimization effect of species hypothesis on statistics in two environments. At the same time, the inference speed of the improved model is about a quarter faster than the FPS of the original Faster-RCNN during inference. Through Qpest-RCNN, the number of insects captured on sticky traps can be counted quickly and accurately and the bivariate kernel density estimation is used to draw the pest density distribution map to observe the pest distribution of the whole farmland visually. This pest-density farmland detection system is valuable for agriculture by helping farmers control pests, reduce crop damage, and increase yield and quality. [ABSTRACT FROM AUTHOR]