关键词: 冬小麦, 无人机, 多光谱, 特征融合, 倒伏

Abstract:

【Objective】The wheat lodging seriously affected the process of photosynthesis and maturity, which led to the reduction of yield and quality. In order to obtain lodging information quickly and accurately, UAV ability of remotely monitoring wheat lodging was evaluated, and a wheat lodging monitoring model was built, so as to provide a technical support for disaster assessment, insurance claims and post-disaster remediation.【Method】Five original multispectral band images, including red, green, blue, red-edge and near-infrared, were acquired by near-ground UAV. The wheat canopy image with flying height of 50 m was preprocessed to obtain digital orthophoto map (DOM) and digital surface model (DSM) with a resolution of 1.85 (cm/pixel), three types of feature information were extracted, namely spectral features, height features, and texture features. Support Vector Machine (SVM) and Random Forest (RF) were used to compare the lodging classification of six different feature set combinations, and accuracy (Acc), precision (Pre), recall (Re) and harmonic mean (F1) were used to determine the best feature combination and classifier. At the same time, three different feature set screening methods (Lasso algorithm, random forest recursive algorithm RF-RFE and Boruta algorithm) were used to comprehensively evaluate the optimized feature subset, and to establish an appropriate evaluation method for lodging classification.【Result】The results showed that the single feature spectrum and texture as well as their combinations had poor classification and evaluation results for wheat lodging, and the “salt and pepper phenomenon” was serious. On this basis, the classification accuracy of DSM information fusion was significantly improved. The random forest classifier was used to combine spectral features, texture features and height features, and the classification accuracy of wheat lodging identification reached 91.48%. In order to reduce the number of feature set variables, three feature optimization methods were adopted. Compared with the full feature set, Lasso algorithm and the RF-RFE algorithm, the optimized feature subset based on the Boruta algorithm had higher classification accuracy and better overall stability. From the perspective of the mean value of the three feature combinations containing DSM, the overall classification accuracy and Kappa coefficient were improved by 0.17% and 0.01 (full feature set), 2.45% and 0.05 (Lasso), 2.87% and 0.05 (RF-RFE), respectively. Among them, spectrum-texture-DSM was the best, with the overall classification accuracy of 92.82% and Kappa coefficient of 0.86.【Conclusion】The study showed that the Boruta algorithm could effectively optimize the number of feature subsets of the spectrum-texture-DSM combination, allow fewer feature parameters to participate in the classification, and obtain higher classification accuracy. Meanwhile, a multi-feature combination-Boruta-RFC technology model was established for accurately monitoring wheat lodging, which provided a reference for wheat disaster assessment and the formulation of remediation measures.

Key words: winter wheat, UAV, multispectral, feature fusion, lodging