Temporal Detection of Pesticide Residues in Tea Leaves Using Hyperspectral Sensing

Tea is considered as a healthy beverage due its antioxidant properties and resultant beneficial effects on human health. India is the second largest producer of tea in the world after China, but every year Tea plants are attacked by several pests and diseases that are responsible for 7-10% loss of the crop. Tea growers spray pesticides on the plants for better control on the pests and to reduce crop loss. The uncontrolled usage of pesticide with ineffective management practices results in high residue levels that are hazardous to human health. In this paper, an attempt has been made to detect the pesticide residue on the tea leaves using hyperspectral sensing. The data on leaves treated with three banned chemicals (Acetamiprid, Cypermethrin and Monocrotophos) and control healthy leaves was collected using spectroradiometer in the spectral range of 350-1052 nm in 213 narrow contiguous bands. The spectral observations of the leaf samples (total of 389 samples) in control and treated with each chemical on consecutive days (i.e. from 1 to 7 days) has been collected. The data in 213 narrow contiguous bands is used as feature set for hyperspectral data analysis. The band selection methods are used to overcome collinearity in spectral observations due to contiguous data. Band selection has been performed by 1. Visual Analysis, 2. Principal Component Analysis and 3. Random Forest based feature importance method. The performance of various classifiers such as Support Vector Machine, Random Forest and Linear Discriminant Analysis has been evaluated using different feature sets and classification scenarios. The classification scenarios considered for each pesticide are 1. two class classification with control and pesticide treated samples, 2. three class classification using control samples, 1 and 2 days after pesticide application, 3. four class classification using control samples, 1, 2 and 3 days after pesticide application and 4. eight class classification with control samples and 1-7 days after application. The PCA based feature selection method was found effective among all. The two class classification scenario of control vs treated provided the best classification performance by using the SVM classifier (accuracy of 98.29%) for all the three pesticides (89.04-98.29%). The three class classification results showed that separation of pesticides with one and two days after application is possible with accuracy of 79.81-95.8%.