A Precision Monitoring Method for Corn Leaf Mite Damage Based on UAV Multispectral Data

Abstract

The corn leaf mite (CLM) is one of the most destructive pests in corn seed production, seriously affecting the yield and quality of corn seeds. UAV-based remote sensing data offers a promising approach for efficiently monitoring CLM under field conditions. However, its performance may be affected by the mechanisms by which CLM piercing-sucking pests cause damage and by field conditions. In contrast, UAV-based fusion of spectral and textural data can improve monitoring accuracy by capturing complementary information from the internal and external aspects of corn leaves. However, there is still a lack of a method for optimizing texture features to achieve precise and efficient monitoring of CLM. This study introduces an index-combination method based on texture indices (TIs) to extract key spatial information from UAV multispectral data for early monitoring of CLM. Initially, a total of 18 vegetation indices (VIs) were extracted from multispectral images captured by UAVs, along with the simultaneous construction of three texture indices derived from textural characteristics. Following this, the Otsu-CIgreen algorithm was employed to determine the best threshold for removing the complex background of the image. Finally, based on the screened VIs and the fused features of VIs and TIs, three machine learning models were used to construct a CLM monitoring model. Among these models, the back propagation neural network (BPNN) model that integrates VIs and TIs performed best, achieving an accuracy of 93.47% and an F1 score of 93.64% on the test set. The results indicate: (1) The BPNN model based on the fusion of VIs and TIs is most effective for CLM monitoring, compared with VIs alone, the fusion of VIs and TIs significantly improves the accuracy of CLM monitoring; (2) Considering the different importance of texture features (TFs) and TIs, the VIs were fused separately with TFs and TIs. The BPNN model based on exponential fusion achieved the highest monitoring accuracy (accuracy: 93.47%, F1: 93.64%); compared with feature fusion, the test set showed improvements (accuracy: 90.00%, F1: 89.67%). The inclusion of texture indices significantly improved the sensitivity and accuracy of early CLM stress detection, providing an effective remote sensing technical pathway for early monitoring and precise prevention and control of CLM, which aids decision-making in smart agriculture and reduces pesticide use.