Vegetation information recognition in visible band based on UAV images

Nowadays, UAV (unmanned aerial vehicle) remote sensing has been widely used in various research fields, due to its incomparable advantages over traditional satellite remote sensing, such as lower cost, fast image access, and high spatial resolution, and so on. But most of the vegetation indices are constructed based on visible bands and near-infrared bands of satellite remote sensing images, and few of them are constructed only based on visible-light bands. Thus, it is necessary to construct a universal vegetation index that is suitable for the visible-light bands of UAV images. According to the analysis of the spectral characteristics of 6 kinds of typical features based on regions of interest in visible-light images from UAV images, this paper proposed a new vegetation index based on red, green and blue bands, named Excess green-red-blue difference index (EGRBDI). The formula of EGRBDI was that the sum value between the square of 2 times green band and the product of red and blue bands divided the difference value of them. The value range of EGRBDI was the interval -1, 1. To determine the accuracy and reliability of EGRBDI, 18 kinds of vegetation indices had been studied in this paper, such as CIVE GLI, ExG, and so on. The overlap between different object types was obtained by calculating the mean value and 1-fold standard deviation of vegetation indices. The results showed that EGRBDI, GLI, ExG, g, CIVE, RGBVI, and V-MSAVI had no overlap between vegetation and non-vegetation information, while other vegetation indices appeared the different degree of overlap. Moreover, EGRBDI effectively enlarged the identification range of vegetation information and reduced the identification range of non-vegetation information. When the grey histogram of vegetation index existed distinct bimodal peaks, the corresponding discrimination performance of ground features was relatively strong. Therefore, the quantized interval of gray histograms should be normalized to the interval 0, 255 for the comparative analysis between the indices. Results of the analysis concluded that EGRBDI, GLI, ExG, g, CIVE, RGBVI, and V-MSAVI had distinct bimodal-peak characteristics and scarcely appeared thorn peaks in the histogram, but the others had either no obvious bimodal peaks or obvious thorn peaks. To determine the thresholds of vegetation information identification, the bimodal histogram method and the maximum entropy method were used to determine the threshold of each vegetation index and got the optimal threshold of each vegetation index by the precision comparison method. The accuracy evaluation results revealed that GBRI and ExB obtained higher classification accuracy by the maximum entropy method than the bimodal histogram method. WI and VEG had the same accuracy between the two methods, and the other 15 indices did better on the bimodal histogram method. Therefore, the maximum entropy method was used to determine the thresholds of GBRI and ExB, while the other indices used the bimodal histogram method to determine their thresholds in this paper. Through the comparative analysis of the experimental results, it could be found that EGRBDI was generally better than the other 18 algorithms and had a great advantage in the case of the low vegetation coverage, which had a total accuracy of 97.67% and a Kappa coefficient of 0.9415. Another 3 UAV images had been used to extract vegetation information of top 5 higher precision indices to further verify the suitability in the area of various ground subjects and used 400 random points to evaluate the vegetation extraction accuracy. The accuracy of the vegetation and non-vegetation information was not less than 90%. The total accuracy in the 3 study areas was higher than 93%. Additionally, the Kappa coefficient was greater than 0.85. The results showed that EGRBDI had been less affected by ground subjects and shadows, and it had better applicability, reliability, and accuracy of vegetation extraction.