Forest Change Detection Method Based On Multi-Source Remote Sensing Images And Deep Learning Techniques

As a carrier of forest resources,forestland is an extremely important national land resource in China.Due to natural or human factors,changes in forestland resources are ubiquitous,which presents challenges for large-scale forestland management.The use of remote sensing technology can effectively improve the efficiency of forestland change investigation,but traditional forestland change detection methods require high data quality and have low detection accuracy.Deep learning is an emerging remote sensing interpretation method,and models based on deep learning technology can achieve higher forestland change detection accuracy,which is of great significance for reducing the workload of related departments.In this paper,an algorithm suitable for forestland change detection is constructed based on multi-source remote sensing images and deep learning technology.First,the design paradigm of the Siamese network forestland change detection model is analyzed and screened to obtain more accurate change features from the front and back temporal images.Second,the SiamFPN-Swin model,which is oriented towards high-resolution remote sensing images,is proposed to achieve higher accuracy in forestland change detection.Finally,to address the problem of traditional forestland change detection methods relying too much on low cloud optical images,the DSNUNet model,which combines optical and SAR images,is proposed,and this model shows good pseudochange suppression capabilities in cloudy areas.The main research conclusions of this paper are as follows:(1)The Siamese network method,as a design paradigm for change detection models,can effectively improve the accuracy of forestland change detection.Compared with the early fusion method,the Siamese network method using the Feature Pyramid Network(FPN)model as the decoder part achieved higher forestland change detection accuracy.When the model used a concatenation information fusion strategy,the Siamese network method achieved the best precision,recall,F1 score,and intersection over union(IoU)indicators,reaching 0.7934,0.7044,0.6866,and 0.5950,respectively.(2)The high-precision forestland change detection model SiamFPN-Swin proposed in this paper for high-resolution remote sensing data achieved the best results in all evaluation indicators,with precision,recall,F1 score,and IoU of 0.8093,0.7767,0.7588,and 0.6723,respectively.Compared with the suboptimal model,the F1 score of SiamFPN-Swin increased by 0.0594.The visualization results show that this model can effectively deal with multi-scale and irregular shape features in forestland change areas,and the detection results have more accurate boundaries.(3)The DSNUNet model,which combines optical and SAR images,can effectively suppress pseudo-changes in cloudy areas and achieve the best precision,recall,and F1 score indicators of 0.6666,0.6626,and 0.6646,respectively.Compared with the suboptimal model,the F1 score of DSNUNet increased by 0.0975.In addition,double Siamese encoder with width differences is proposed,which can extract accurate change features from different data sources with information differences.When the initial channel numbers of the optical and SAR branches are 32 and 8,respectively,the performance of DSNUNet is optimal.