Location: Crop Production Systems ResearchTitle: Using machine learning and hyperspectral images to assess damages to corn plant caused by glyphosate and to evaluate recoverability
|ZHANG, TONG - Hangzhou Dianzi University|
|YANG, PINGTING - Hangzhou Dianzi University|
|ZHANG, JINGCHENG - Hangzhou Dianzi University|
Submitted to: Agronomy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/15/2021
Publication Date: 3/19/2021
Citation: Zhang, T., Huang, Y., Reddy, K.N., Yang, P., Zhang, J. 2021. Using machine learning and hyperspectral images to assess damages to corn plant caused by glyphosate and to evaluate recoverability. Agronomy. 11:583. https://doi.org/10.3390/agronomy11030583.
Interpretive Summary: Widely used glyphosate has been known for off-target crop damage and resulting in yield losses. A rapid approach for assessing crop damage is needed in comparison with traditional laborious and time-consuming approaches. Scientists from Hangzhou Danzi University and USDA-ARS Crop Production Systems Research Unit at Stoneville, MS have collaboratively developed a hyperspectral imaging method to assess the plant damages over corn sample from the field. With the analysis of machine learning algorithms this study indicated that the spectral response of sprayed glyphosate to corn damage can be captured and the newly developed method can accurately differentiate the recoverable and unrecoverable corn damages. The results have practical significance for improved weed management.
Technical Abstract: Glyphosate is most widely used herbicide due in part to widespread adoption of glyphosate-resistant crops and has a potential to drift onto non-target crops from ground or aerial applications, resulting in severe yield losses. Therefore, the research on glyphosate is of great practical significance for sustainable agricultural production. Based on the advanced hyperspectral imaging (HSI) technology and field experiment, the spectral responses of corn plants under different glyphosate concentrations were measured, and the corresponding characteristic bands were screened and the model was constructed to evaluate the degree of glyphosate injury to the crop. Plants were treated with glyphosate at 6 concentrations to simulate injury. It was found that 1 week after treatment (WAT), 2 WAT and 3 WAT showed orderly spectral changes. When the concentration of glyphosate was equal or greater than 0.5X (X = 0.87 kg ae/ha representing the recommended use rate for GR corn), corn was damaged irreparably. On the basis of selected three best bands, two recoverability spectral indices (crop damage ratio index and crop damage normalized index) were created. Based on the Jeffries-Matusita distance and successive projections algorithm, the band sensitivity on glyphosate spray concentration was analyzed. Using three machine learning algorithms of k-nearest neighbors, random forest and support vector machine, the quantitative model of crop damage with glyphosate spray concentrations was created. The spectral response of glyphosate sprays to corn damage can be clearly captured by the HSI technology. The recoverability spectral indices can accurately differentiate the recoverable and unrecoverable damages, and the overall accuracy (OA) is higher than 95%. Without distinguishing between the recoverable and unrecoverable damages, the best spectral feature set was determined to assess the effect of the concentrations of glyphosate spray. In this case the highest overall classification accuracy was 58%. Only when the crop can be recovered, the classification accuracy is relatively high (OA = 75%). These results have practical significance for weed management.