| Accurately obtaining the spatial distribution and temporal changes of grape planting areas is crucial for refined planning and management,as well as high-quality base construction.Currently,remote sensing images are commonly used for the segmentation and change detection of large-scale planting areas,but the dispersed spatial positions of grape planting areas and complex background environments can reduce the accuracy of these methods.With the development of deep learning,semantic segmentation networks that can automatically extract image features have become an effective way to achieve high-precision segmentation and change detection of grape planting areas.However,many existing segmentation and change detection models in the field of image processing have overlooked the features of remote sensing images,which have multiple bands and varying scales of ground object types.To address this issue,this article comprehensively explores the characteristics of grape planting areas and constructs a segmentation model for grape planting areas with enhanced band information,as well as a change detection model for grape planting areas with multiscale information fusion.These improvements enhance the recognition accuracy of spatial distribution and temporal change information in grape planting areas.Finally,the paper presents a segmentation and change detection system for grape planting areas,which can promote research in related fields.The main contributions of this work include:(1)A segmentation method for grape planting areas based on band information enhancement is proposed in this study.Segmentation of grape planting areas in remote sensing images is challenging due to the complex types of ground objects and the similarity of textures between some ground objects and grape planting areas,leading to low segmentation accuracy.To address this problem,we utilize the Deep Labv3+ network to improve the number of input channels,allowing us to receive more spectral information.We also construct a band information enhancement module and use the correlation between feature maps of each band to generate comprehensive features.The proposed method,called BIE-Deep Labv3+,is trained on the 2016 and 2019 Gaofen 2 image grape planting area datasets and tested on the 2020 image.Our results demonstrate that the BIE-Deep Labv3+ model output has significantly improved segmentation accuracy,with mean pixel accuracy,mean intersection over union,and frequency weighted intersection over union of 98.58%,90.27%,and 97.34%,respectively.Compared to the Deep Labv3+ model,the BIE-Deep Labv3+ model improved by 0.38,2.01,and 0.63 percentage points,respectively.The BIE-Deep Labv3+ model has a larger receptive field,captures multi-scale information features,and amplifies the differences between ground objects,thereby solving the problems of texture similarity between classes,complex backgrounds,and environmental factors in the image of grape planting areas.The predicted grape planting areas are more complete while reducing model parameters,and the edge recognition effect is good.This provides an effective method for segmenting grape planting areas in remote sensing images with complex backgrounds in larger areas.(2)This study proposes a method for detecting changes in grape planting areas based on multi-scale information fusion.The current challenges in change detection after classification are error accumulation and a decrease in the representation ability of small-scale features in the high-level feature maps of deep learning models.To address these challenges,we use Resne Xt as the encoder and Deep Lab V3+ as the decoder to construct a siam neural network change detection model.In the combined network,we use atrous convolution for downsampling to reduce the loss of image features caused by pooling operations.We also replace the last layer of ordinary convolution with atrous convolution to increase the receptive field of the model.Additionally,by adding an attention mechanism before the feature map enters the decoder Deep Lab V3+,we overcome the heterogeneity problem of information fusion in both spatial and channel directions.This enables us to fuse high and low layer information in the network and fully consider feature information at different scales.The proposed ADM-Resne Xt Deep Lab V3+ model is evaluated on the 2016,2019,and 2020 Gaofen 2 image grape planting area change detection datasets.Our results show that the mean pixel accuracy,mean intersection over union,and frequency weighted intersection over union for grape planting area change detection results are 98.89%,83.99%,and 97.93%,respectively,which are 0.13,2.07,and 0.26 percentage points higher than the basic model.The model effectively improves the issues of false positives and missed detections by providing relatively complete edges and interiors of the change areas identified by the model.It has generalization performance.Our proposed method provides a good idea for detecting changes in grape planting areas with dispersed distribution and uneven areas.(3)A segmentation and change detection system for grape planting areas is proposed.To meet the application requirements of grape planting area segmentation and change detection in related research,a client-free B/S system architecture was adopted.The system was implemented using Spring Boot and Vue framework coding for the server and web ends,respectively.The system can provide users with information related to segmentation and change detection models and enable them to perform segmentation and change detection functions for grape planting areas in uploaded images.The system has been tested and meets established business requirements while providing users with an efficient and flexible interactive experience.The proposed scheme provides a way to obtain spatial location information and temporal change information of grape planting areas,and the segmentation and change detection system provides methodological support for extracting information from grape planting areas in complex background images. |
PDF Full Text Request