11:00 AM - 1:00 PM
[MTT46-P02] Search for Machine Learning Suitable for Crop State Discrimination by Angle-Dependent Multispectrum
Keywords:Machine learning, Angle dependent spectra
In remote sensing using microsatellites and drones, which is indispensable for the next generation of agriculture, multi-wavelength spectral imaging in a narrow band is expected to improve the discrimination accuracy, but it is not widely used at present. If the full spectrum is captured each time, the cost will not be reduced, although the proof of principle can be achieved due to 1) a huge amount of data 2) limited field of view because the spectrum changes depending on the angle of incident and detected light, and 3) still expensive infrared detectors. The spectra observed from crops have few characteristic spectral peaks, as seen in the field of materials science. This also makes it difficult to discriminate them because the spectral changes used for state discrimination are very small. In order to overcome these problems, it is important to construct a spectral library, which is a detailed database of the relationship between the light irradiation angle, measurement direction, elevation angle, and the spectra on the ground, and to combine it with data on crop conditions and yields.
In this study, we compared both Support Vector Machine (SVM) and Convolutional Neural Network (CNN) methods in order to find an appropriate machine learning method to select narrow bands from angle-dependent spectra that are characteristic for state discrimination. A total of 3675 sets of angle-dependent spectra of the leaf color scale (CF360 for paddy rice, Fujihira Industries), which is used as an indicator of fertilizer management in paddy rice, were taken indoors using a halogen lamp in the spectrum range of 450-700 nm. The classification accuracy was compared between SVM and CNN. The narrow-band wavelengths of 5, 10, and 30 nm were used for comparison.
The results showed that the classification accuracy of SVM was high and that the accuracy was high when the measurement elevation angle was 38-54 degrees. Based on these results, we measured the actual spectra of Chinese chives and daffodils to prevent accidental ingestion of both and discussed the extraction of narrow bands by SVM and its classification accuracy.
In this study, we compared both Support Vector Machine (SVM) and Convolutional Neural Network (CNN) methods in order to find an appropriate machine learning method to select narrow bands from angle-dependent spectra that are characteristic for state discrimination. A total of 3675 sets of angle-dependent spectra of the leaf color scale (CF360 for paddy rice, Fujihira Industries), which is used as an indicator of fertilizer management in paddy rice, were taken indoors using a halogen lamp in the spectrum range of 450-700 nm. The classification accuracy was compared between SVM and CNN. The narrow-band wavelengths of 5, 10, and 30 nm were used for comparison.
The results showed that the classification accuracy of SVM was high and that the accuracy was high when the measurement elevation angle was 38-54 degrees. Based on these results, we measured the actual spectra of Chinese chives and daffodils to prevent accidental ingestion of both and discussed the extraction of narrow bands by SVM and its classification accuracy.