Deep Learning Based Coastal Feature Extraction And Intelligent Interpretation

Coastal zone is the interface between the land and the ocean.Due to the continuous interaction between the ocean and the land,coastal zone is the most active natural area on the earth’s surface.Meanwhile,it’s also an area with the most superior resources and environmental conditions.Currently,more than 50% of the world’s population lives in the coastal zone.Obviously,coastal zone is of great significance in economic,environmental,and military.Over the past decades,the increase of population and the rapid urbanization process have heightened tensions in human-environment interactions which makes coastal zone vulnerable to climate change and natural disasters.To address these issues,the Land–ocean Interactions in the Coastal Zone(LOICZ)and Future Earth-Coast(FEC)projects were established to develop the capacity to assess,model and predict change in the global coastal zone under multiple forcings(including human activity)and consequences for human development.Taking Guangdong-Hong Kong-Marco Greater Bay Area as the main research area,based on deep learning methodology,this thesis takes coastal zone feature extraction,coastal zone utilization suitability evaluation and coastal zone land use and land cover change(LUCC)prediction research content through summary and review of related researches.In view of questions exist in current research,several new deep learning based methods are proposed and effectiveness of these methods are verified by multiple experimental scenarios.The specific research work of this thesis is as follows:(1)Currently,deep learning based coastal zone feature extraction methods are mainly applied research of vanilla deep learning model and there is a lack of adaptive improvement according to the characteristics of coastal zone.As a result,existing methods have room for improvement in accuracy,computational speed,and dependence on perfectly annotated samples.In this regard,this thesis selects three problems of coastline extraction,satellite-derived bathymetry and coastal land use classification as the research objects,and proposes corresponding intelligent extraction methods.Specifically: multi-scale coastline extraction model based on quadtree decomposition and deep learning,filtering out inland regions containing a large number of confusing objects through quadtree decomposition and realizing quick and accurate coastlines extraction on large-scale imagery.Extraction experiments were conducted using data from the Greater Bay Area,with a total of 117,374 tile images.The results indicate that,compared with direct semantic segmentation,the proposed method can effectively filter the confusing object and resulting in an accuracy improvement of about 6%.In addition,from the result of proposed method,more than 90% of the result area are confirmed at TMS level 14 or above,in other word,number of tile images to be processed is greatly reduced so that required computation time is significantly reduced.In this way,this thesis extracts the coastline of the Greater Bay Area with a resolution of 2 m,which provides accurate data for subsequent satellite-derived bathymetry,land use classification and suitability evaluation.Deep convolutional network based satellite-derived bathymetry model,mining semantic information through convolution network structure and non-value part of ground truth label is processed through Masked Loss so as to improve the accuracy and stability of bathymetry model.Yongle Atoll scenario and the Greater Bay Area scenario are constructed for experiments,and the SVR model is used as a control.Area of Yongle Atoll scenario is small,thus factors of remote sensing imagery are consistent,whereas,water depth points are of uneven distribution,more than 90% of the points are within2 m.The Greater Bay Area scenario consists of 6 Sentinel-2 images and acquisition condition of these images are of internal differences.The results show that in the Yongle Atoll scenario,all deep learning method can more accurately retrieve water depth in the extremely shallow area,which increases the total RMSE by 0.1.Similarly,in the Greater Bay Area scenario,deep learning models are adaptive to remote sensing imagery with different acquisition condition and water bodies with different turbidity.From the deep learning result,gradual transition from nearshore to open water depth is clear,and the MAE is generally within 10 m.On the contrary,the SVR model is seriously affected by acquisition condition and bathymetry accuracy is generally lower than that of deep learning,and MAE in some areas are exceeds 20 m.Coastal land use extraction model based on electronic map data and deep learning method,road network and base map information provided by the electronic map are used to divide boundary of functional areas,followed by attribute assignment though a classifier combining shape features,POI attribute and image features,by this means,high accuracy land use results are extracted in the absence of perfectly annotated label.In this thesis,the land use extraction experiment is carried out in Huizhou,and outputs result with a resolution of 2m.The experimental results show that classifier based on POI or scene classification only can only achieve 90% accuracy in some categories,while the fusion method proposed in this thesis can achieve 90% accuracy in all categories.(2)Based on the extraction of coastal elements,the suitability of coastal construction land is evaluated.Current models are mainly based on evaluation indicators from multiple fields,calculate suitability value corresponding to a single evaluation indicator,and weights these elements to obtain the final result.Considering characteristics of the coast of Greater Bay Area,storm surge risk evaluation result is taken as one of the assessment indicators.Due to current models utilize pixel with limited information as processing unit,it’s difficult to support complex classifiers,so that information loss is likely to occur when processing evaluation indicators from multiple fields with huge internal differences.To solve the problem,coastal construction suitability evaluation model based on deep learning is proposed,which can mine spatial feature of data through the convolution structure and fuse features from different evaluation indexes through multi branch network.The experimental results indicate that the proposed model can generate more complete results and can also produce reasonable progressive results in inshore areas,in addition,the model has better generality.With the extraction results obtained in the previous sections,this thesis constructs two experimental scenarios,one contains storm surge data,the other does not.The results indicate in the scenario with storm surge data,the proposed method can produce reasonable progressive results in inshore areas,as a contrast,the result of Land USEM,a typical suitability evaluation model,is of discrimination.In the scenario without storm surge data,the proposed method can also produce continuous and complete result.(3)Current coastal LUCC prediction models mainly calculate prediction results pixel by pixel through simple transition rules.Such structure limits the prediction accuracy and shape and detail of prediction results are also different from real data.such structure to some extent ignores the heterogeneity of local regions,which has been demonstrated to be important for urban development.To address this problem,a GAN-based coastal zone land use classification prediction model is proposed to mine local multi-scale spatial information from heterogeneity of local regions,which has been demonstrated to be important for coastal city development.The Attention mechanism is also introduced to improve the effect of prediction simulation.The experimental results demonstrated that the proposed model achieved a higher accuracy,and the evaluation using the local landscape pattern index also indicated that the shape and detail produced by the proposed method were also closer to the real data.Consequently,based on the deep learning,this thesis takes the Guangdong-Hong KongMacao Greater Bay Area as the main research area to extract coastal elements,evaluate the suitability of construction land,and conduct LUCC prediction.The experimental results confirm that deep learning methods have excellent performance on these issues.As a result,the deep learning method provides a new solution for recognizing the natural laws,evaluating the resources and predicting the changes of the coastal zone.