Research On Remote Sensing Image Classification Based On Causal Inference And Neural Network

Remote sensing image classification is a key technology in geographic information science,environmental monitoring,and urban planning.By accurately identifying and segmenting surface features of the earth,remote sensing image classification can promote understanding and analysis of the geographic environment.However,complex scenes pose challenges for traditional remote sensing image classification methods.In recent years,deep learning techniques have provided new solutions for remote sensing image classification,but their generalization ability and interpretability still need to be improved.This paper first analyzes the challenges and problems in remote sensing image classification and proposes a causal inference and neural network-based remote sensing image classification method aimed at improving the generalization performance of remote sensing image classification models.Finally,the effectiveness of the proposed method is validated through experiments on two publicly available remote sensing image datasets.The main research contents of this paper are as follows:1)Firstly,causal inference theory is introduced into the remote sensing image classification task,and a neural network method based on causal probability graphical models,called Causal Neural Network(CANN),is proposed.The network structure of CANN includes separation module,intervention module,and loss module,effectively solving the confounding factors and spurious correlations present in remote sensing image semantic segmentation,thereby improving the accuracy and robustness of semantic segmentation.2)Secondly,the counterfactual content interpretation method in causal inference is improved to make it suitable for semantic segmentation scenarios.Considering the characteristics of remote sensing image classification models and the requirements for interpretation,this paper proposes a method to construct a fine-grained concept library using ontology triplets to better support remote sensing image classification tasks.Using the SLIC algorithm to automatically generate superpixel segmentation and assign concept labels to pixel sets,the training set of the concept library is expanded.Human-understandable high-level concepts generated by counterfactual content explanation can be used to optimize the training process of the model.3)Then,a technical route combining causal inference for remote sensing image classification is proposed,which guides the deployment of appropriate data augmentation methods to optimize the performance of the remote sensing image classification model based on counterfactual interpretation results.By analyzing the counterfactual explanation content,the shortcomings of the model in predicting certain categories can be better understood,and targeted data augmentation schemes such as sample balance and transformation augmentation can be used to produce enhanced samples to solve these problems.4)Finally,to verify the effectiveness of the proposed method,this paper applies it to two publicly available remote sensing image datasets,Love DA and Vaihingen,for experiments,revealing that the remote sensing image classification method based on causal inference and neural networks can effectively improve the generalization ability and interpretability of remote sensing image classification models under biased data.In comparison with other mainstream methods,the proposed method is proved to be effective and superior in remote sensing image classification tasks.The experimental results show that the proposed method achieves significant performance improvement on these two datasets compared with the baseline model.This indicates that the proposed causal inference neural network method can effectively eliminate dataset bias and improve the generalization ability of the model.