Remote sensing monitoring of areca yellow leaf disease based on UAV multi-spectral images

Yellow leaf disease is a serious disease that endangers the growth of areca, it is urgent to monitor the infection severity and spatial distribution in time and accurately. However, the traditional monitoring methods are mainly depend on visual inspection and manual investigation, which affects the efficiency and spatial scope of monitoring. Low altitude UAV remote sensing technology can effectively solve the problems of insufficient optical satellite images acquisition caused by cloudy and rainy weather in areca planting area, and improve the real-time monitoring of areca yellow leaf disease. In this paper, in order to identify the severities and spatial distribution of areca yellow leaf disease, five band(including blue, green, red, near-infrared and red-edge wavebands) multispectral images were obtained by using the MicaSense RedEdge-M multispectral camera mounted on the DJI Phantom 4 Pro V2.0. Based on the Minimum Redundancy Maximum Relevance (mRMR), three sensitive features were selected from 15 potential vegetation indexes. Using Back Propagation Neural Network(BPNN), Random Forest(RF) and Support Vector Machine(SVM) classification algorithms respectively, a monitoring model of areca yellow leaf disease severity was constructed. A total sixty in-situ sampling points were selected and the disease index (DI) were obtained, according to the characteristics of the disease and the separability of remote sensing images, the severities of the disease were divided into three grades: health (DI<1%), slight (1%≤DI＜10%) and serious (DI≥10%), and the number of samples was 18, 22 and 20 respectively. According to the priority principle of importance of feature variables, Ratio Vegetation Index (RVI), Modified Simple Ratio Index (MSR) and Anthocyanin Reflectance Index (ARI) were finally selected. Two-tier neural networks including hidden layer and output layer were used to build the BPNN model. The results showed that the overall accuracy (OA) of BPNN model was 91.7%, which was 6.7% and 10.0% higher than that of RF model and SVM model, respectively. The Kappa coefficient of the BPNN model was 0.875, which was the highest among the three models. In general, the omission errors and commission errors of BPNN model were the smallest, the errors of health, slight and serous levels were 11.1%, 15.8%, 13.6%, 9.5% and 0, 0 respectively. Consequently, it is feasible to monitor the severities of arecanut yellow leaf disease based on the UAV multispectral image. The study can provide a reference for the diseases monitoring of other tropical crops.