Crop classification recognition based on time-series images from HJ satellite

Abstract: Time-series satellite images can reflect the seasonal variation from vegetation on land surface, and have better performance than single-temporal image for vegetation classification. Multi-temporal satellite images such as Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Very High Resolution Radiometer (AVHRR) have been widely used for crop classification throughout the growth season, but exhibit some limitations due to lower spatial resolution. On the other hand, some satellite imagery data with medium- resolution (such Landsat TM) and high-resolution (such QuickBird) also display some weaknesses thanks to lower temporal resolution. Environment Satellites HJ-1A/B of China have a better spatial resolution of 30 m than MODIS and AVHRR, and a higher temporal resolution of 2 days. So it is noticeable to use the time-series images from HJ satellites for crop classification.In this paper, selecting the largest farm, Youyi Farm in Nongken Region, Heilongjiang Province, China as an example, ten HJ-CCD time-series images from June to September 2010 were used to classify crops in the farm. After atmospheric and geometric corrections, SPLINE algorithm was applied to remove cloud in images for reconstructing time-series images. By collecting three main crops (soybean, rice and corn) ground truth data with Global Positioning Systems (GPS) in fields, the band reflectance of Red and NIR, and vegetation indices of NDVI and EVI with temporal changes were extracted. The red band reflectance of rice between in June 2nd to July 12th and August 26th to September 1st had significant difference between rice with others crops. The EVI of corn was less than soybean from July 12th to September 1st. After analyzing the images through serial threshold division, masking treatment, assisting with background data and expert knowledge, the decision tree classified arithmetic was established. Then, support vector machine (SVM) and maximum likelihood supervised classification method were also used to identify these crops.The results indicated that HJ-1A/B satellite had a particular advantage in extracting vegetation information with its higher spatial and temporal resolutions. Cloud processing was of importance to reconstruct no cloud time series data. According to temporal changes of spectral reflectance and typical vegetation indices of different crop ground samples, all crops had similar tendency of NDVI. So NDVI was difficult to identify different crops. Both the red band reflectance and EVI had the remarkable spectral features to reflect the different crops growing and vegetation coverage information. Growing individual, isolated crops in bulk has become common for large-scale farms in Heilongjiang Nongken region. Planting information of soybean, corn and rice were successfully extracted based on the time series images by three methods. Comparing SVM and maximum likelihood supervised classification method with decision tree classified arithmetic, the results suggested that decision tree classified arithmetic could effectively achieve the accurate classification of main crops, its overall accuracy reached up to 96.33%. Different growth may have the similar variation tendency and so be confusion. While time series images can clearly show different spectral feature curve in different crop growth stage, avoiding wrong or missing category and greatly improving classification accuracy.