Identification of pests in cotton field based on hyperspectral data

Objective  To identify cotton pests quickly and accurately without destruction, and formulate pertinently a plant protection spraying plan.

Method  Hyperspectral data of cotton leaves were collected and analyzed. FieldSpec^®3 portable spectrum analyzer with a wavelength range of 350−2 500 nm was used to obtain hyperspectral data of cotton leaves including normal leaves and leaves infected by aphids and red spiders.K-nearest neighbor and SVM algorithm were used to distinguish above leaves. In order to further optimize pest identification of the model and improve the recognition accuracy, the principal component analysis method (PCA) was used for feature dimension reduction, and the grid search method was used for parameter optimization.

Result  The models of pest identification were constructed by K-nearest neighbor algorithm and SVM algorithm, and recognition rates of two models were 86.08% and 89.29% respectively. Recognition rate increased after introducing PCA for feature dimension reduction and using grid search for parameter optimization. The recognition accuracies of K-nearest neighbor algorithm and SVM algorithm reached 88.24% and 92.16% respectively.

Conclusion  Hyperspectral data can be used to distinguish aphid or red spider-infected leaves and normal cotton leaves. Using PCA dimensionality reduction and grid search method, the recognition rate can increase without obtaining specific characteristic bands. For identifying aphid- or red spider-infected leaves and normal leaves, SVM algorithm based on radial basis kernel function is better than K-nearest neighbor algorithm.