Land-cover Semantic Change Detection Method Of High Resolution Optical Remote Sensing Images

As an important basic task in remote sensing image interpretation,the main purpose of remote sensing image change detection is to process and analyze image data,and extract pixels or regions in multi-temporal images where the land-cover category changes.Currently,in the actual production of land-cover classification and change detection data,manual interpretation methods are often used to visually interpret images,which is time-consuming and cost high labor.A new generation of deep learning theoretical methods have become research hotspots,although the current change detection method based on the fully convolutional network model in deep learning has achieved better performance than traditional change detection methods,a series of problems still exist,for example,low feature extraction and model generalization capabilities for multi-source remote sensing images and insufficient training sample data.In addition,most of the existing methods only study the change extraction under a single land-cover category and the binary change detection under multiple land-cover types.It is difficult to know the specific land-cover change information under the two temporal images,and also difficult to directly convert it into actual production results.The semantic change detection derived from binary change detection can further obtain the land-cover category information of the two-temporal image and clarify the specific category change information,which is more in line with the needs of actual production.In response to the above challenges,this thesis proposes an integrated network model based on a deep fully convolutional network for the land-cover semantic change detection task that simultaneously extracts the land-cover category information and the change detection area.The main tasks are as follows:(1)As for land-cover classification tasks,this thesis designed a compound dilation convolution module,a feature recalibration module based on cascaded attention mechanism,a feature pyramid structure based on cross-level connections and multiple combined loss functions,and constructed an FPSeg Net model.As for binary change detection tasks,this thesis designed dual-input FPBcd Net model in data-fusion and siamese structure based on the basic modules and overall structure of the FPSeg Net model.Aiming at the integrated extraction task of land-cover classification and change detection of two-temporal remote sensing images,this thesis constructs a dual-input and three-output FPScd Net integrated model.It further optimizes the general structure of the model,reduces model parameters and designs training goals and training strategies in order to realize integrated extraction of semantic change detection.(2)This thesis collects remote sensing imagery data under multiple time phases in typical domestic areas,designs the construction process of sample datasets in combination with manual sample labeling and constructs the Fujian dataset with seven land-cover categories and the Henan dataset wtih five categories.These datasets can simultaneously meet the requirements of model training and test accuracy evaluation for single-phase land-cover classification task,two-phase binary change detection task and semantic change detection task,which can help alleviate the lack of public largescale dataset in the current field to a certain extent.Three tasks of single-phase landcover classification,two-phase binary change detection and semantic change detection were carried out on the two constructed semantic change detection datasets.Through the construction of sample datasets and comparative experimental analysis,the effectiveness and practicability of this method are verified.