基于无人机遥感多光谱影像的棉花倒伏信息提取

为在棉花发生倒伏灾害后快速获取田块尺度下的受灾信息,该文以2017年8月21日强风暴雨导致大面积棉花倒伏的新疆生产建设兵团第八师135团的部分田块作为研究区,由无人机遥感试验获取倒伏后的多光谱影像,通过分析倒伏和正常棉花的光谱反射率差异提取了多种植被指数和主成分纹理特征,结合地面调查样本建立了3种花铃期倒伏棉花的Logistic二分类模型并进行了精度评价和验证。结果表明:棉花倒伏前后在可见光波段的反射率差异微小,而在红边和近红外波段的反射率明显降低0.12～0.20;以第一主成分均值(PCA1_mean)建立的Logistic二分类纹理模型效果最优,在测试集上分类结果的准确率为91.30%,ROC(receiver operating characteristic)曲线距左上角点最近,AUC(area under the roc curve)值为0.80。通过将该模型应用于试验区影像,分类制图效果良好且符合棉田倒伏症状特点。该研究可为无人机多光谱遥感棉花灾损评估提供参考。

Information extraction of cotton lodging based on multi-spectral image from UAV remote sensing

Extracting crop lodging information, such as spatial location and area, is very critical to agricultural disaster assessment and agricultural insurance claim. It is hard work to measure the lodging information using traditional methods such as a ground survey. A survey method using remote sensing techniques can quickly and efficiently obtain crop lodging information, but it is limited by the lack of timely and available satellite remote sensing data. In recent years, the application of unmanned aerial vehicles (UAV) develop rapidly in the agricultural field, which makes UAV equipped with image sensors become a portable, stable and efficient crop survey tool with the characteristics of low cost, high timeliness, and small weather impact. A few scholars have measured the lodging area of wheat and corn crops using visible or multispectral images. However, studies using UAV multispectral images to survey cotton lodging information have not been published. Therefore, a survey method of cotton lodging using multi-spectral image was derived from UAV remote sensing experiment which was carried out in the 135th Regiment of the 8th Division of Xinjiang Production and Construction Corps on August 23 of 2017. In this study, the spectral characteristics of lodging and normal cotton were first analyzed and summarized, and a series of vegetation indices were calculated. 16 texture features of the first two components were calculated according to gray level co-occurrence matrix (GLCM) after principal component analysis (PCA), and the optimal texture features were selected in terms of the coefficient of variation (CV) and the relative difference (RD). The result showed that it was apparently different between lodging and normal cotton in spectral curves and texture features. Compared with normal cotton, the difference in reflectance of the lodging cotton in the visible wavebands was small, while was significant in the red and near-infrared bands, in which the reflectance dropped about 0.12-0.20. The main reason for this phenomenon might be the collapse of the cotton canopy structure. Mean of the first principal component(PCA1_mean), PCA1_entropy, PCA1_homogeneity, PCA2_mean, and PCA2_homogeneity texture features had the lower CV and higher RD, which were very suitable for classification of normal and lodging cotton. Then, 10 vegetation indices and 5 texture features of the measured samples were calculated as characteristics index, and the training set and test set were divided. Forward stepwise was used to select the best features on the data set. Binary Logistic models on lodging and normal cotton classification were constructed with different features combination, including spectral model, texture model, and spectral-texture model. The prediction accuracies of the classification models were evaluated by ground survey samples. All classification models had a good classification effect on lodging and normal cotton. Among them, the texture model constructed with the PCA1_mean had the highest precision, and the classification accuracy on the test set was 91.30%. The classification accuracies of spectral-texture model and spectral model were following, but the classification accuracy was also more than 85%. Finally, the classification models were applied to the multi-spectral image at the pixel level, and 3 thematic classification maps were created. Compared with the visual interpretation results, the texture model has the best classification effect. The "salt-and-pepper plaque" of the thematic map was the least, and the lodging crop had the characteristic of aggregation occurring in space. The ROC(receiver operating characteristic) curve was closest to the upper left corner and the calculated AUC(area under the ROC curve) value was 0.80. According to the results of the study, we may safely draw the conclusion that the method to extract lodging cotton information using the multi-spectral image of UAV remote sensing based on optimum texture features is accurate. The lodging classification has a high accuracy of mapping, which is basically consistent with the actual lodging in the field.