Geographical principles of remote sensing image analysis and the hierarchical patch model based analysis framework

In the past two decades, Geographical Object-Based Image Analysis (GEOBIA) has been widely studied and applied; however, it still does not meet the expectation for big remote sensing image analysis in geographical cognition activities in terms of accuracy and intelligence. We think that the major problem is the lack of geographical thoughts to lead the research and development (R&D) of GEOBIA key techniques, especially when introducing the techniques of computer vision, which does not regard comprehending the earth’s surface as the objective. On this basis, we review the concepts of GEOBIA from a geographical perspective, specifically the principles of region, scale, and pattern and function. From the region principle, we regard the image segmentation in GEOBIA grouping the spatial neighbor pixels sharing similar spectral and textures as the representation of a fine-scale geographical zoning in remote sensing image spaces. From the scale principle, we regard the multiscale of segmentation as the representation model quantifying the relationship of geographical zones among different scales. From the pattern and function principle, we regard the multiscale segmentation as an ideal hierarchical patch model representing the earth surface structure, i.e., the landscape, and could quantify the pattern (e.g., orientation, shape, arrangement, distance, etc.) for the function recognition. In other words, we think that the target of GEOBIA is to recover the hierarchical multiscale structure of the earth’s surface from the remote sensing images so that we can quantify the structure and then recognize and comprehend its function. On the basis of these reviews, we propose an iterative GEOBIA framework where the core is constructing a hierarchical patch model of the earth’s surface. The framework starts with fusing the big geographical data, including remote sensing images, existing geographical thematic maps, and other helpful knowledge, to construct an initial hierarchical patch model of the earth’s surface. Then, object features are extracted from the hierarchical patch model, and the function of these objects is recognized; the features include the internal features extracted from the object itself (e.g., shape and spectrum) and the external features extracted from its relationship with other objects (e.g., its neighbor objects, parent objects, and children objects). Finally, the recognized results are used to update the hierarchical patch model for the next recognition cycles. With the iteration of the remote sensing image analysis, the accuracy of geographical object recognition can be improved because we also have a more accurate hierarchical patch model describing the earth’s surface due to the updating process, which could provide an accurate calculation of the object’s features. To achieve the above proposed framework, we also propose a few suggestions for further R&D, such as constructing different hierarchical patch models for different geographical elements, fusing multiresolution images by using the hierarchical patch model instead of pixels, and choosing the suitable interpretation models instead of one model for different big geographical patches. We hope the above insights could provide an instructive idea of how to embed geographical knowledge into intelligent interpretation models to extract new knowledge from big remote sensing images with improved accuracy. © 2024 Science Press. All rights reserved.