Early season detection of rice plants using RGB, NIR-G-B and multispectral images from unmanned aerial vehicle (UAV)

Crop plant detection is vital for mapping crop planting area and extracting pure crop canopy information. In this study, three cameras (RGB, color infrared (NIR-G-B) and multispectral (MS) camera) were mounted on a multirotor unmanned aerial vehicle (UAV) to obtain images of rice canopy at the early growth stages (tillering, jointing and initial booting stages). We proposed a new decision tree (DT) combining texture features (mean and variance (C. V)) and spectral features (TS-DT) for rice plants detection within UAV images. First, the image was classified into the pure class and the mixed class based on the C. V value. Then the pure class was classified into rice plants and road by the DN or reflectance value in red band. The mixed class was classified into rice plants and background (soil, water and duckweed) through comparing each pixel value to the mean value within the moving window. The results showed that TS-DT exhibited an averaged high classification accuracy with overall accuracy (OA) and kappa coefficient (KC) of 91.25%, 0.86, 92.88%, 0.86 and 93.53%, 0.88 for RGB, NIR and MS image among the early three growth stages, respectively. The highest estimation accuracy was obtained at booting stage and the lowest was at tillering stage. Compared with the traditional classification methods, the TS-DT method achieved an improved estimation accuracy of 5.2-26.71% in OA and 0.06-0.40 in KC. Therefore, this TS-DT method is a reliable approach for crop plants detection using UAV imagery.