Building feature learning method for incremental data of remote sensing image

Abstract

The invention discloses a building feature learning method for incremental data of remote sensing image, which comprises the following steps of: screening initial stock data and constructing a typical sample library, and generating and constructing a typical feature library corresponding to the typical sample library based on initial model parameters; training the newly-added data samples and typical sample library samples together to acquire optimization model parameters, and estimating the losses of new and old data through a cross entropy method and a feature matching method; after training is completed, selecting new data training samples according to a certain criterion to update and replace the typical sample library and the feature library; and continuously optimizing and iterating building feature extraction parameters of the model along with continuous increase of new data. According to the method, model parameters can be efficiently optimized, the adaptability to old data characteristics is maintained while the method adapts to new data characteristics, and the effect similar to the effect of retraining the model by mixing all new and old samples is achieved on the premise that the training time is greatly shortened.

Description

technical field [0001] The invention relates to the field of surveying and mapping science and technology, in particular to a building feature learning method for remote sensing image incremental data. Background technique [0002] In practical applications, remote sensing image data is often an incremental process of dynamic increase and continuous accumulation. With the increase of data, when the new and old data are heterogeneous due to different imaging areas or imaging conditions, it is difficult to achieve good results by directly applying the model trained by stock data samples on incremental data; but if new data samples are used for The parameters of the original model are adjusted and the new model is obtained. Its accuracy performance on the new data is significantly improved, but it is also accompanied by performance degradation on the old data, that is, the model "learns the new and forgets the old". This forgetting problem can be improved by isolating new and ...

AI Technical Summary

Problems solved by technology

[0003] Incremental learning research on remote sensing image data has been reported as early as the end of the last century, but it has not attracted much attention in the field of remote sensing for a long time; in recent years, scholars have gradually carried out research in the field, but few have paid attention to single-category detection of buildings Research on the cognitive performance of the model on new and old data. The existing building feature incremental learning model still mainly solves the knowledge forgetting problem caused by new heterogeneous data through parameter isolation. However, the large volume and continuous growth of high-resolution remote sensing images characteristics, may lead to the infinite growth of model parameters with the data, which will bring heavy storage burden and model maintenance costs. There is still a lack of efficient incremental learning methods with relatively fixed model parameters in the field.

Therefore, how to build a reasonable incremental learning framework to achieve efficient model parameter optimization in terms of shortening training time and limiting parameter volume, there is currently no clear solution

Images