Small sample city remote sensing image information extraction method based on meta learning and attention

Abstract

The invention provides a small sample city remote sensing image information extraction method based on meta-learning and attention, which comprises the following steps: constructing a small sample city remote sensing information pre-training model, and in a pre-training stage, performing pre-training network learning on a small sample set to fully learn feature information of existing samples, so as to obtain a small sample city remote sensing image information pre-training model; obtaining an initial feature parameter of the small sample set and a deep convolutional network trunk; constructing a'time-space spectral angle 'attention model for enabling the network to pay attention to important'time-space spectral angle' information in the training process, suppressing noise and redundant information and improving the classification performance of the model; and establishing a meta-learning and attention cooperation mechanism and realizing small sample city remote sensing information extraction, including performing parallel correlation learning on a feature parameter obtained by a ''spatio-temporal spectrum angle'' attention model and an initial feature parameter, introducing a regularizer to minimize cross entropy and structural risk, and realizing full mining of small sample information. By applying the method, the precision of extracting the typical ground feature information of the small sample city is higher.

Description

technical field [0001] The invention belongs to the field of remote sensing image information extraction, and designs a small-sample urban remote sensing image information extraction method based on meta-learning and attention coordination mechanism. Background technique [0002] Cities are important places for human survival, activities, and production. With the rapid development of urbanization, "urban diseases" such as ecological function degradation, environmental pollution, and frequent extreme weather events seriously threaten the sustainable development of cities. With the rapid development of remote sensing technology, especially domestic remote sensing satellite technology, China already has the data conditions to independently obtain global high-resolution urban geographic information. The extraction of urban features from images still faces great challenges. [0003] The deep learning network has strong nonlinear representation ability and high image recognition ...

AI Technical Summary

Problems solved by technology

With the rapid development of remote sensing technology, especially domestic remote sensing satellite technology, China already has the data conditions to independently obtain global high-resolution urban geographic information. The extraction of urban features from images still faces great challenges

[0003] The deep learning network has strong nonlinear representation ability and high image recognition ability, and has obvious advantages in high-resolution image information extraction, but the deep learning algorithm usually requires more parameters and a large-scale image knowledge base. Restricted by factors such as geographical location and weather, it is difficult to label large-scale and large-scale training samples, and the labor cost is high. For some urban areas with small samples, deep learning algorithms are prone to weak generalization capabilities, which is not conducive to obtaining accurate training samples. Urban Remote Sensing Information

Images