Detection of Aconitum leucostomum based on a ResNet deep residual network

The spread of poisonous grasses under the influence of human activities and climate change threatens the security of ecosystems and healthy development of animal husbandry. Aconitum leucostomum is one of the most serious poisonous grasses in the Yili region of Xinjiang. To achieve the goal of rapid, accurate, and automatic detection of poisonous grasses in natural grassland under a heterogeneous background, A. leucostomum was used as the research object, and a data set of this grass was constructed by aerial orthographic imagery by UAV. Based on the Faster-RCNN and SSD algorithms, ResNet50 and ResNet101 trunk network features were used to extract features and the detection accuracy of the different methods was compared. A comparison of the detection accuracy showed that the test set Faster-RCNN_ResNet50 had the highest mAP (mean average precision) value of 64.74% while SSD_ResNet50 had the lowest mAP of 48.70%. The mAP values of Faster-RCNN_ResNet101 and SSD_ResNet101 were 63.37% and 52.55%, respectively. This study has reference significance for the detection of A. leucostomum from aerial orthophoto images.