Evaluating the Potential of Sentinel-2 for Low Severity Mites Infestation Detection in Grapes
| Autor: | Srinivasu Pappula, Jayantrao Mohite, Navin K. C. Twarakavi |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
010504 meteorology & atmospheric sciences
biology 0211 other engineering and technologies Hyperspectral imaging Feature selection 02 engineering and technology biology.organism_classification medicine.disease_cause 01 natural sciences Random forest Spectroradiometer Lasso (statistics) Infestation Mite medicine Image resolution 021101 geological & geomatics engineering 0105 earth and related environmental sciences Mathematics Remote sensing |
| Zdroj: | IGARSS |
| Popis: | The Mite is one of the major sucking pests in grape which goes undetected in its initial phase as the symptoms are not easily visible to the naked eyes. In this paper, we address the problem of mites infestation detection using temporal hyperspectral data and also evaluate the potential of using Sentinel-2 data for mites infestation detection. The reflectance data from grape leaves with healthy and low infestations of mites have been collected using spectroradiometer. The hyperspectral remote sensing data is collected from 213 bands with wavelength ranging from 350 nm to 1052 nm during 15th Jan – 18th Feb 2017. Variations observed in the spectral reflectance over time makes the detection based on multitemporal data difficult. Data in 213 narrow contiguous bands is used as feature set for hyperspectral data analysis but this large feature set may cause the over-fitting problem and also poses the requirement of large storage and greater processing time. To avoid this, feature selection using Least Absolute Shrinkage and Selection Operator (LASSO) has been carried out to get the optimum band set. Features selected by LASSO were fed to classifiers such as Random Forest (RF), Artificial Neural Network (ANN) and Logistic Regression (LR) to evaluate their performance. Results suggest that LR based model provides maximum accuracy of 93.24%. In addition to this, to investigate the potential of using Sentinel-2, data in 213 narrow bands were simulated to Sentinel-2. Data has been simulated to 10 and 20m spatial resolution bands available in 350–1050nm range. This simulated 8 band feature set has been fed to the same set of classifiers to evaluate their performance. Results suggests that LR provides maximum classification accuracy of 89.12% using simulated Sentinel-2 bands. Further to validate the algorithm using actual ground observations from the field, we have implemented simulated Sentinel-2 based algorithm on two Sentinel-2 images available during the study period and results are compared with actual ground observations about mites infestation. Results suggest mites detection accuracy of 83.33% which shows the good agreement and potential of Sentinel-2 for mites infestation detection. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|