Graduation project - BIT-based change detection of bitemporal remote sensing images (Python development)
Click me to download the file
Thesis directory
Compressed package directory
Graduation project
Summary of content (main problems and difficulties to be solved):
Grasping the utilization rate of land resources and land cover types is an important content of geographic national conditions census and monitoring. Efficiently obtaining accurate and objective land use conditions can provide decision-making support for national and local geographic information. With the development of remote sensing and sensor technology, especially the popularization of multi-temporal high-resolution remote sensing image data, we can grasp the subtle changes of any surface in the world without leaving home. Change detection focuses on the change detection task of buildings, that is, given the bitemporal remote sensing images of a certain area, it is required to obtain the building changes in this area.
Main task:
Improve the BIT model and apply it to solve the remote sensing image change detection task. The main work includes: data enhancement of the original data set to improve the generalization ability of the training model; resampling of samples to solve the problem of sample imbalance; adding weight factors when calculating cross entropy loss, so that the model pays more attention to changes Loss of samples and indistinguishable samples; use orthogonal experiments to combine different hyperparameters, and perform range analysis on the obtained results to determine the factors that have the greatest impact on the model and the best hyperparameter combination scheme; in the original BIT The hollow convolution pooling pyramid module (ASPP) is added to the backbone network to extract multi-scale image features, and improve the model's missed detection and rough boundaries of multi-scale targets; at the same time, a dual attention mechanism (CBAM) is added to ResBlock for In order to improve the model's attention to changing pixels; finally, by comparing the AC-BIT model with other models, it can be concluded from the experimental results that the comprehensive effect of the AC-BIT model is better than other models.
Remote sensing image change detection refers to determining the different processing processes of the same object or phenomenon based on observations at different times. The purpose of its research is to find out the interesting change information and filter out the irrelevant change information as interference factors. The results of change detection can be used for land resource utilization, land cover type monitoring, crop extraction and crop planting area acquisition, efficient acquisition of accurate and objective land use conditions, and monitoring of land changes. With the development of remote sensing and sensor technology, especially the popularization of multi-temporal high-resolution remote sensing image data, the field of remote sensing in my country has entered the fast lane of high-resolution images, and the demand for analysis and application services of remote sensing data is also increasing day by day. Traditional methods have poor ability to characterize high-resolution satellite remote sensing images and rely on manual experience and a huge workload. With the rise of artificial intelligence technology, especially the image recognition method based on deep learning has achieved great development, and related technologies have also promoted the transformation of the field of remote sensing. Compared with the traditional visual interpretation method based on human sea tactics, remote sensing image recognition technology based on deep learning can automatically analyze the types of objects in the image, showing great potential in terms of accuracy and efficiency.
This graduation project improves the BIT model and applies it to solve the remote sensing image change detection task. The main work includes: data enhancement of the original data set to improve the generalization ability of the training model; resampling of samples to solve the problem of sample imbalance; adding weight factors when calculating cross entropy loss, so that the model pays more attention to changes Loss of samples and indistinguishable samples; use orthogonal experiments to combine different hyperparameters, and perform range analysis on the obtained results to determine the factors that have the greatest impact on the model and the best hyperparameter combination scheme; in the original BIT The hollow convolution pooling pyramid module (ASPP) is added to the backbone network to extract multi-scale image features, and improve the model's missed detection and rough boundaries of multi-scale targets; at the same time, a dual attention mechanism (CBAM) is added to ResBlock for In order to improve the model's attention to changing pixels; finally, by comparing the AC-BIT model with other models, it can be concluded from the experimental results that the comprehensive effect of the AC-BIT model is better than other models.
Mastering the utilization rate of land resources and land cover type is an important content of geographic national conditions survey and monitoring. Efficiently obtaining accurate and objective land use conditions can provide decision-making support for national and local geographic information. With the development of remote sensing and sensor technology, especially the popularization of multi-temporal high-resolution remote sensing image data, we can grasp the subtle changes of any surface in the world without leaving home. Change detection focuses on the change detection task of buildings, that is, given the bitemporal remote sensing images of a certain area, it is required to obtain the building changes in this area.