Global Vision Media Focus
The company was featured on CCTV a total of 15 times, and its reach through Party media, central media, and local official media reached a total of 50 million people.
Exploration of Automatic Land Parcel Extraction Methods Based on Satellite Remote Sensing Imagery
Background Introduction
Plot extraction is a crucial foundational task in remote sensing interpretation, land resource management, and agricultural spatial information technology. Plot extraction refers to identifying the spatial extent of individual plots from satellite imagery, UAV imagery, or other remote sensing data, forming areal objects with clearly defined boundaries. For agricultural production management, accurately obtaining plot boundaries is essential for farmland registration, crop monitoring, yield estimation, land use surveys, and the distribution of agricultural subsidies. Traditional plot extraction methods often rely on manual visual interpretation or traditional image processing methods such as edge detection and clustering segmentation. These methods are effective in areas with regular plots, clear images, and significant differences in land features, but they still have significant limitations when dealing with complex agricultural scenarios.
For example, the size of plots varies greatly in different regions, and some small farmland plots have fragmented and irregular boundaries; there are also significant differences between different seasons, different sensors, and images with different resolutions; at the same time, factors such as roads, ditches, field ridges, and vegetation shadows can also interfere with boundary recognition. Therefore, traditional methods struggle to maintain stable accuracy under large-scale, multi-scene conditions.
In recent years, the development of deep learning technology has provided new approaches for automatic land parcel extraction. Related research has transformed farmland boundary extraction from the traditional "boundary pixel recognition" task to "land parcel instance segmentation," meaning it no longer simply determines whether a pixel belongs to the boundary, but directly identifies the complete extent of each independent land parcel. This approach better preserves the integrity of land parcel objects and reduces problems such as adjacent parcels overlapping, boundary breaks, and missed detections. Existing research has also constructed datasets containing multi-resolution satellite imagery and a large number of annotated land parcel instances, with image resolutions ranging from approximately 0.25 meters to 10 meters, capable of adapting to the needs of farmland parcel extraction at different scales.
Model Structure
The land parcel extraction model is an end-to-end automated extraction model for satellite remote sensing imagery. It consists of six parts: satellite imagery, preprocessing, land parcel extraction model, instance results to vector, merging of the same vector results, and vector output, as shown in Figure 1.
Figure 1.Model processing flow
First, the model uses satellite imagery as input data, leveraging information on features such as farmland, roads, ditches, and field ridges to provide a foundation for parcel boundary identification. Then, the input imagery undergoes preprocessing, including image cropping, coordinate registration, format standardization, normalization, and image slicing. Based on this, the parcel extraction model performs feature analysis and instance segmentation on the preprocessed imagery, distinguishing adjacent parcels to avoid overlap or missegmentation between different parcels. The instance results output by the model need to be further transformed...Vector dataThis involves extracting the outline of land parcels from a mask and generating closed polygons for use in GIS spatial analysis. Since large-scale satellite imagery is typically segmented into multiple image patches for processing, the same land parcel may be repeatedly identified or segmented in different slices. Therefore, it is necessary to merge vector results belonging to the same land parcel to eliminate duplicate boundaries and maintain the integrity of the land parcel morphology. Finally, the model outputs complete land parcel vector results, which can be saved in common spatial data formats such as Shapefile and GeoJSON for use in land parcel statistics, area calculation, etc.Land use analysisAnd spatial planning and management.
Results Display
Figure 2 shows the vector output of the model. The model can accurately identify regular farmland, large contiguous farmland, and some small-scale fragmented farmland. For areas with clear boundaries and regular plot arrangements, the extraction results basically match the actual plot boundaries. For areas with densely packed plots and complex shapes, the model can still maintain good plot separation performance, reducing the possibility of adjacent plots being mistakenly merged. The plot extraction method based on deep learning instance segmentation can effectively improve the automation level and extraction accuracy of farmland boundary identification, providing important technical support for agricultural remote sensing monitoring and refined land resource management.
Figure 2.Vector results