A tactile recognition method for rice plant lodging based on adaptive decision boundary

Rice plant lodging recognition is key to improving the performance of weed control components. However, there are several challenges that cause existing methods to have insufficient capability to handle this task: (1) the complexity of the rice field environment, including disturbed backgrounds and rice plant leaf canopy overlay and (2) the large number of actual samples that cannot be fully encompassed by the model inputs. In this study, an adaptive decision boundary-based tactile recognition method for rice plant lodging is proposed to address the two aforementioned problems. We used the homemade rice plant lodging recognition device to obtain rice plant tactile data, formed the energy-clustering coefficient (ECC) feature vectors using the data splicing method, and constructed a rice plant lodging recognition model using the adaptive decision boundary method. The field test of rice plant lodging recognition indicated that the Accuracy and F1-score of rice plant lodging recognition were 91.80% and 91.82%, respectively. The results of the performance comparison test of operational parameters suggested that the accuracy of rice plant lodging recognition increased with the increase in traveling speed and weeding wheel speed, but the stability of recognition decreased. The results of the study are important for improving the quality of weed control in rice fields.