结合HJ卫星影像和最小二乘孪生支持向量机的小麦蚜虫遥感监测

Remote sensing monitoring of wheat aphids by combining HJ satellite images with least squares twin support vector machine model

为了准确、及时地监测小麦蚜虫发生情况,利用野外定位调查数据及环境与灾害监测预报小卫星星座HJCCD和HJ-IRS影像数据,在北京市通州区和顺义区小麦蚜虫发生的关键生育期(灌浆期),提取对蚜虫病情影响较大的小麦长势因子和生境因子,利用最小二乘孪生支持向量机建立该研究区的小麦蚜虫监测模型,并与传统支持向量机、费歇尔线性判别分析和学习矢量量化神经网络模型的监测结果进行对比。结果表明:最小二乘孪生支持向量机模型的总体监测精度达到86.4%,优于传统支持向量机模型(77.3%)、费歇尔线性判别分析模型(77.3%)和学习矢量量化神经网络模型(72.7%),取得了较好的监测效果。

Summary Pests and diseases have become serious because of global warming, which have caused great economic losses to agricultural production, and have threatened human life and health, so it was very urgent and challenging to prevent or control pests and diseases. Real-time dynamic monitoring of the occurrence of pests and diseases in large scale continuous space can guide the prevention or control work accurately and effectively to reduce the impact of pests and diseases as well as the environmental pollution caused by the indiscriminate use of pesticides. Remote sensing technology can provide effective information for crop pests and diseases monitoring quickly and accurately on a massive continuous spatial surface. HJ-1A/1B satellite has a high revisit period （4 days）. Multi-spectral images obtained by HJ-1A/1B satellite sensors have high spatial resolution （30 m） and are very suitable for the monitoring of agricultural pests and diseases. The occurrence of wheat aphids affects seriously the yield and quality of wheat. Monitoring of the wheat aphids accurately and timely is helpful for effective prevention and control of pests. In this paper, by using the field location survey data and the HJ-CCD and HJ-IRS image data, the growth factors and the environmental factors of wheat are extracted, including normalized difference vegetation index （NDVI）, green normalized difference vegetation index （GNDVI）, reflectance of red band, land surface temperature （I.ST） and perpendicular drought index （PDI）. These factors had a great influence on the occurrence of wheat aphids. The monitoring model of wheat aphids in Tongzhou District and Shunyi District of Beijing was established by using the least squares twin support vector machine （LSTSVM）. The LSTSVM has a good processing ability for large scale unbalanced data and has stronger robustness than the traditional support vector machine （SVM）. Computational complexity of LSTSVM is reduced by using the least squares algorithm to transform inequality constraints into equality constraints. Experimental results showed that： the overall monitoring accuracy of the LSTSVM model was 86.4% and the Kappa coefficient was 0.71; the traditional SVM model was 77.3% and 0.52; the Fisher linear discriminant analysis （FLDA） model was 77.3% and 0.54; and the learning vector quantization （LVQ） neural network model was 72.7% and 0.39. In sum, the algorithm proposed in tiffs paper has higher precision than the traditional SVM, FLDA and LVQ neural network.