Search

Public Earth Observation (EO) data archives, e.g., MODIS, Landsat, and Sentinels, are valuable sources of information for a broad range of applications. For decision-supporting applications used in urban planning, land management, and sustainable development, images covering regions similar to the study area are prerequisites for high-accuracy decision making. These desirable images cannot be quickly searched for in the EO data archives via image metadata alone but can be obtained through content-based image retrieval methods. Land cover (LC) information, traditionally obtained through image segmentation or classification processing, is typically used in existing methods. Image processing is time consuming and has various accuracy levels for heterogeneous images, thus decreasing retrieval efficiency and accuracy. Additionally, the monotemporal LC information used has a limited ability to distinguish among confusable regions with different terrain, e.g., forests located on flatlands or mountains, and to obtain regions, e.g., urban regions, with similar growth rates. In this study, we employ free multiple-year 30 m LC products, a terrain product, and the Google Earth Engine (GEE) platform to accurately and efficiently locate the desired heterogeneous moderate spatial resolution images from various public EO data archives. Regions similar to the query region are detected with two-stage similarity calculations: First, monotemporal pixel-based LC and terrain information are used to filter out the most dissimilar regions; second, object-based LC change and terrain information are used to locate similar regions. Then, the desired images covering these detected similar regions are obtained from EO data archives via image metadata, e.g., geographical location and acquisition time. The experimental results of the two representative query regions show that our method can be used to obtain the desired images within several minutes and has higher accuracy than the LandEx method and a simplified method using only monotemporal LC information. The main contribution of our study is to reveal that LC changes and terrain information are helpful for improving the retrieval accuracy achieved from monotemporal LC information alone. Our method has great operability, with no need to perform EO data acquisition, image processing of raw EO images, or management of computational resources. Our method is conducive to making full use of images in various public EO archives to improve the decision making quality of decision-supporting applications.