Abstract: Abstract: Identification of paddy fields in the hilly regions is important for policy making of food self-sufficiency in China. However, extracting image information using current image analysis techniques is difficult because of the unique terrain of hilly regions. The traditional pixel-based analysis of remotely sensed data is usually affected by pixel heterogeneity, mixed pixels, and spectral similarity, thus leading to the inaccurate identification of paddy fields in hilly regions. This study aimed to find other methods for accurate paddy field identification in hilly regions. The study area was Xiangtan City located in the mid-east of Hunan province, a good representative of hilly regions. In Xiangtan city, the land use change markedly increases with rapid economic development, leading to gradual decline of cultivated land. The Chinese environment and disaster mitigation satellite (i.e., HJ-1A/1B) image of the region was data source for land use map. The HJ-1A star was equipped with a charge-coupled device (CCD) camera and a hyperspectral imager, whereas the HJ-1B star was equipped with CCD and infrared (IR) cameras. The satellite observes the ground in widths of 700 km with a ground pixel resolution of 30 m by four multispectral imaging steps. The object-oriented image analysis technique is a new type of automatic technique under a computer environment. The information carrier used was multi-scale objects composed of multiple adjacent pixels containing rich semantic information. Image segmentation is an important classification step because high-precision remote sensing (RS) image classification depends on good segmentation. The multi-scale image segmentation algorithm was applied in the preliminary object extraction to fully interpret the RS images with the different spectral features, shape, and textural features of real ground targets. The configuration of multi-scale segmentation thresholds directly affected the integrity of features extracted from RS images. In this study, the cultivated and uncultivated lands were segmented with the scale of 40; then the cultivated land was further segmented under the scale of 30 and 20, respectively. By comparing and analyzing the segmentation results on the two scales, the optimal scales for the extraction of paddy fields in different regions were configured selectively. The phenomenon of different objects with the same spectral characteristics and same object showing different spectral characteristics may occur in the classification of RS images. The two phenomena pose challenges for RS image interpretation. In order to identify the information related with paddy field distribution in hilly regions, the key point is the RS identification between paddy field, dry field, forest and grassland. According to the classification features, k-nearest neighbor (KNN) classifier and decision tree classifier were employed to interpret the RS images of paddy field in hilly regions. The KNN classifier was improved by dividing the training samples into three sets. The result of the improved KNN classifier was better than that of traditional methods. The precision of the improved KNN classifier was 74.6%. However, the total precision and Kappa coefficient of the decision tree classifier were higher than the KNN classifier. The total identification precision of the former reached 90.25%, with commission error rate of 4.12%, omission error rate of 5.63%, and Kappa coefficient of 0.79. A comparison of the results of the two classifiers showed that the decision tree classifier is more suitable for paddy field identification based on object-oriented analysis in hilly regions.