Paddy planting acreage estimation in city level based on UAV images and object-oriented classification method

Abstract: Unmanned aerial vehicle (UAV) is a highly efficient technology to detect land surface, which is capable of acquiring centimeter-level spatial resolution data and acquiring ground survey samples information timely and accurately. This is the basis for large-scale crop acreage estimation. However, the usual procedure of UAV image for sampling plot identification is still hand-generated, which leads to time-consuming and high cost expense. Therefore, it is urgent to develop an efficient approach to extract crop acreage in sampling plot from UAV orthographic image to support the regional crop acreage estimation. In this paper, we introduced an object-oriented automatic classification method, instead of the visually digitized method, to identify UAV quadrat paddy, and the combination of the quadrat paddy data and the paddy classification result from satellite remote sensing was exploited to estimate the paddy planting acreage. The comprehensive comparison between the methods of manual visual interpretation and object-oriented classification was carried out. This experiment was conducted in Pinghu City, Zhejiang Province. The integrated satellite imageries of Chinese GF-1 WFV sensor and American Landsat8 OLI were acquired. According to paddy phenology calendar of the study area, there were 2 key phases for the paddy identification. In middle June, paddy was in the seedling growing season, which represented the water spectrum. However, in later July, the paddy rice field showed the vegetation spectral information in satellite imagery, which was at tillering stage. The information sources of GF-1 WFV and Landsat8 OLI acquired at the respective phenological stages were sufficient to map paddy rice distribution. First, the data were preprocessed for geometric correction, and atmospheric correction was applied for both satellite images. Support vector machine (SVM) was used to identify the vegetation and water components from GF-1 data and Landsat8 data, respectively. Then, a logical "and" operation was conducted between vegetation and water to generate the paddy rice spatial distribution. UAV images were obtained from T10 Bumblebee platform, and a total of 7 sampling belts were acquired. Amount of UAV photos was firstly tiled by Pix4D mapper to generate UAV images with the resolution of 0.08 m. Then, UAV images were segmented by multi-scale algorithm in eCognition Developer, the segmentation scale was set to 50, 100, 150, 200 and 250 and the spectral standard deviation of image objects and its rate were calculated at each scale to determine the optimal segmentation scale. After various tests, the choice of 200 was decided as the optimal segmentation scale to describe the field boundary clearly and correctly. Then, different features were constructed and nearest neighbor classification method was adopted to extract the paddy planting distribution in sampling belts. A framework of 300 m × 300 m square grids was built covering the extent of paddy as the primary sample unit, and then the sampling ratios were calculated using acreage index to allocate the samples in each stratum. Then, the regional paddy rice acreage was estimated by combining extracted paddy rice acreage based on the satellite remote sensing and UAV sampling belt. The results showed that the overall classification accuracy of paddy from UAV image was more than 93% for the object-oriented automatic classification method, which met the basic requirements of building the samples. Most of all, the difference of CVs (coefficient of variations) of the acreage estimation aided by automatic classification quadrat data and manual visual interpretation quadrat data was 0.0008, which stated the object-oriented automatic classification could achieve the same estimation performance as artificial visual interpretation method. That implies it is feasible to apply the object-oriented automatic classification in place of the visual digitization method to extract the quadrat data to support the regional paddy acreage estimation. This achievement can be applied and tested extensively in large-scale areas and with different crops.