Semantic segmentation method based on multi-source heterogeneous remote sensing image

Abstract

The invention discloses a semantic segmentation method based on a multi-source heterogeneous remote sensing image, and relates to the technical field of deep learning. The method specifically comprises the following steps: step 1, preprocessing training pictures in a remote sensing image data set; step 2, establishing a multi-stage segmentation head network, and completing feature extraction and segmentation prediction of the training pictures by using the multi-stage segmentation head network to obtain a segmentation result with multi-stage labels; step 3, performing multi-level label supervised training on the multi-level segmentation head network built in the step 2 to obtain a semantic segmentation model; 4, segmenting the remote sensing image to be segmented; and step 5, obtaining a final segmentation result by fusing prediction results of the multi-stage segmentation heads. The method has the advantages that pixel-level classification is carried out on the given remote sensing image, so that a semantic segmentation result is accurately given.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a semantic segmentation method based on multi-source heterogeneous remote sensing images. Background technique [0002] The classification of surface feature elements based on remote sensing images is a systematic system for classifying relatively fixed objects on the ground surface, and it is one of the important means of surface feature element observation and mapping. It has a very wide range of applications, such as: cultivated land red line prediction, ecological red line prediction, etc. However, due to the influence of phenomena such as the same object with different spectra and the same spectrum with different objects, it is extremely difficult to analyze and process remote sensing images. Extraction method of surface features. [0003] At present, the rapid development of convolutional neural networks has made great progress in semantic segmentation. However, th...

AI Technical Summary

Problems solved by technology

In general, it has the following technical difficulties: 1) The labels of remote sensing image datasets are often multi-level, and how to make good use of this type of labels is an important issue; 2) Remote sensing images usually have extremely high The resolution, such as 5000×5000 pixels, is much larger than conventional semantic segmentation data; 3) affected by the phenomenon of "same object with different spectrum" and "same spectrum with different object", the domain of different remote sensing image datasets The difference is very large, which will cause the accuracy of the model to drop sharply when switching to a new data set; 4) Due to the limitations of application scenarios, the task of classification of ground features usually needs to be completed under limited storage resources and computing resources. Very strict requirements on model design

Images