Research On Heavy Metal Pollution Recognition Based On Deep Learning In Hyperspectral Remote Sensing

Hyperspectral technology brings a new opportunity to identify heavy metals contamination in soil,and has drawn increasing scientific attention in recent years.Due to the information redundancy among bands,the interferences among elements in the soil,and the introduced instrument noise in the process of spectroscopy,the spectrum and the metal content in the soil have a complicated mapping relationship,which makes it difficult to establish an effective prediction model.In addition,due to the different observation scales between indoor spectroscopy and hyperspectral remote sensing,the established inversion model by indoor spectrum is difficult to apply to hyperspectral images.To address the above problems,the following researches are carried out to realize the recognition of heavy metal pollution in soil by hyperspectral images.(1)The inversion model of soil heavy metal based on Hyperspectral is established,including double-layer random forest band selection and inversion model establishment.The double-layer random forest is used to select spectral bands to reduce model overfitting caused by too many spectral features.Otherwise,this method can avoid the model generalization performance degradation caused by empirical band selection.The results show that the twolayers random forest could effectively select the bands related to the soil heavy metal information,improving the prediction accuracy.These conclusions are the basis for identification of soil heavy metal pollution using hyperspectral image in the study area.(2)A deep learning model for hyperspectral image classification is established.The deep learning model is composed of a cascade structure with residual structure,multi-scale spatial feature modules and attention mechanism.The residual structure is easy to optimize and reduce network degradation.Multi-scale spatial feature module is used to extract feature information with different receptive field ranges.The attention mechanism redistributes the weights of the features.In addition,the global average pooling as a transition layer was used to transform high-dimensional feature information into feature vectors.The validation experiments were carried out on three public datasets(i.e.,Indian Pines,Salinas,and Pavia University datasets).Experimental results showed that the proposed model was better for perceiving threshold changes and effectively extract the data features of a smaller proportion of categories.The proposed model improved the classification effect of hyperspectral remote sensing images.(3)The hyperspectral image of the research area captured by the ZY1-02 D satellite,and the geochemical data of Zn,Cu,and As at scale of 1:200000 in the study area are used to construct training data and test data.The proposed model was applied to the identification of heavy metal pollution in the soil.The results showed that the proposed model has high accuracy in identifying heavy metal pollution in this area.