Vegetation indices derived from digital camera images have been shown to be useful for rice growth diagnosis in previous research. However, various image processing methods were proposed for calculation of vegetation indices, and they were not standardized. Hence, we compared several image processing methods to identify the suitable method for evaluation of rice leaf area index and above ground biomass using RGB images and false color images. Vegetation indices were calculated with different image processing method, and prediction accuracy for rice growth from the vegetation indices were evaluated. As a result, color and brightness correction improved relationships between rice growth and vegetation indices but segmenting the rice plant area and reduction of gamma correction were not necessary. Through this study, the findings showed the possibility of rice growth diagnosis using aerial images taken by unmanned aerial vehicles.
This paper presents a crop classification method using synthetic aperture radar (SAR) satellite data for mapping, in place of existing ground surveys. We used TerraSAR-X X-band dual-polarization data and RADARSAT-2 C-band full-polarization data. Values of the sigma-naught and polarimetric parameters were calculated from each type of data. We assessed the accuracy of classification performed by the random forests machine-learning algorithm. Three results were obtained. First, the classification accuracy was evaluated using RADARSAT-2 data for five scenes. Using nine variables calculated from each scene of RADARSAT-2 data, the overall accuracy exceeded 0.92. Second, the classification accuracy was evaluated using both RADARSAT-2 and TerraSAR-X data for five scenes. Using nine types of variables in the RADARSAT-2 data and four types of variables in the TerraSAR-X data, a significantly higher overall accuracy (over 0.93) was obtained than using only RADARSAT-2 data. This demonstrates the advantage of using SAR data for the two types of bands. Finally, for economic efficiency, the number of SAR scenes used for classification was reduced. The classification accuracy using only three scenes of RADARSAT-2 and TerraSAR-X data was not significantly different from that using five scenes. This shows that classification is efficient without requiring a large quantity of data.