Improved forest change detection method for remote sensing imagery using UNet++
Current deep learning-based models used to detect changes in forest cover suffer from two major issues:complex structure and neglect of the spectral and spatial synergistic relationship.These limitations often result in unsatisfactory detection results.To address these challenges,we propose an improved UNet++lightweight forest cover change detection method that combines multi-scale spatial decoupling convolution(MSDConv)and a spatial-spectral feature cooperation strategy(SSFC).Using this method,a non-weight sharing pseudo-twin network was initially constructed based on the UNet++network,which allows for better feature extraction while adding only a small number of parameters.The MSDConv module was adopted to capture the multi-scale features of the changing objects,thereby reducing information redundancy and parameter computation.Subsequently,SSFC was introduced into the MSDConv module to obtain three-dimensional attentional weights between spatial and spectral networks without adding extra parameters.This enables MSDConv to produce richer edge and detail features.Having assembled this framework,six vegetation indices were used to enhance the forest cover change features.Using the proposed model,we obtained values of 93.12%,93.62%,and 93.37%for the accuracy,recall,and F1-score,respectively.Additionally,the number of parameters and computational volume of the model were found to be 6.28 MB and 11.25 GB,respectively.Compared with the original Sami-UNet++method,our proposed model exhibited only slight decreases in accuracy,recall,and F1-score,with values lower by only 1.41%,1.66%,and 1.53%,respectively.However,the number of parameters and computational volume is significantly declined by 5.76 MB and 16.19 GB,respectively.The application of our model offers a significant improvement in the efficiency of detecting changes in forest cover.This advancement proves invaluable when dealing with large amounts of image data,and provides a technical means for the assessment of forest hazards,as well as the protection of forest resources.
国家科技期刊平台
NSTL
万方数据
