Identification Of Typical Targets For Oil Storage And Transportation In Ports By Remote Sensing Images

The imports and exports of oil affect the world’s political and economic balance because oil is a strategic reserve material.The oil strategy of estimating the oil reserves and forecasting the oil development trends around the world has great guidance for the oil procurement of our country.Currently,58% of the global oil trade relies on maritime transportation.The most common storage tool for oil is the storage tank,and to facilitate maritime transportation,large oil reserve based are built in ports.Therefore,our ship transportation and shipping economy are also affected by the oil strategy.The estimation of oil volume in tanks and the identification of oil tankers on the route and port is an effective means of estimating oil reserves and predicting the trend in reserves.To get the object of oil tanks and oil vessels on a global scale,this thesis uses high resolution optical remote sensing images to detect oil tanks and oil tankers.This thesis focuses on the calculation of tank volume based on remote sensing images.Then the object detection algorithm based on the convolutional neural network(CNN)is used to improve the detection accuracy of oil tanks and ships;finally,the fine-grained classification method is used to identify oil tankers.The main innovative works in this thesis are as follows:(1)In this thesis,a calculation method is proposed for the length of crescent-shaped shadow of oil tanks.After extracting the tank shadows,find the points corresponding to the upper and lower edges of the shadow according to the sub-pixel subdivision localization idea,find the shadow lengths corresponding to all pixel points,and then take the median of all shadow lengths as the final shadows length.The tank height can be calculated according to the geometric relationship between the shadow length and the tank height;the tank radius is detected using the hough transform;finally,the tank volume is calculated using the cylinder volume calculation formula.In this thesis,the experiment was conducted using a remote-sensing image of Gaofen 2.The absolute error range between the estimated volume and the actual tank volume is between 0.0416×10~4m~3～0.3050×10~4m~3,and the relative errors ranged from 0.38%to 2.78%,indicating that the method in this thesis has some practical application value.(2)The Hough transform has a missed detection when detecting oil tanks,and an enhanced task aware spatial disentanglement(ETSD)method based on the CNN is proposed in this thesis to improve the recognition ability of oil tanks.The task aware spatial disentanglement(TSD)method finds the feature region misalignment problem for both classification and location regression and uses different deformable pooling operations to obtain the respective suitable feature regions for them.However,two problems of the TSD method are found in this thesis.One is that there are feature conflicts between classification and location regression in the auxiliary branch.The second is that the deformable pooling operation is not suitable for the judgment of negative samples,and for positive samples,the effect gradually decreases with the increase of the Intersection of Union(Io U)between the samples and the real bounding box and even produces negative effects.This thesis first replace the auxiliary branch with double heads to separate the parameters of classification and regression to alleviate the feature conflicts between them;then,we add weights based on the Io U to the loss value of the auxiliary branch to further improve the accuracy of deformable pooled features.In this thesis,experiments are conducted on the two tank datasets from DIOR and DOTA datasets,and the average precision AP values reach 54.43% and 72.08%,respectively,which are 0.49% and 0.43% better than TSD,and outperform many other CNN based object detection algorithms.(3)Although the above method improves the identification ability of tanks,the positioning accuracy decreases compared to the hough variation.To further improve the tank positioning accuracy,this thesis combines the enhanced task aware spatial disentanglement method with the Cascade region convolution neural network(RCNN).Cascade structure allows further positional regression for samples.For the problem of threshold change due to the change of the Io U distribution,this thesis proposes a dynamic Io U threshold method based on Gaussian skew distribution,which uses the mean and skewness values of Intersection of Union within a certain range to calculate the new thresholds,and the method can continuously adjust the thresholds according to the changes of the Intersection of Union distributions.The average precision AP values reached 54.99% and 75.00% on the tank data of DIOR and DOTA datasets,respectively,which improved by 0.56% and 2.92% compared to the ETSD,and outperformed many other CNN based object detection algorithms.The final calculated tank volume error ranges from0.01% to 2.67%.(4)The above method is designed for horizontal bounding boxes and cannot be directly applied to the detection of rotated ships.This thesis firstly proposes a rotation invariant task aware spatial disentanglement(RITSD)method;secondly,proposes an angle deformable pooling model,which can improve the accuracy of the predicted angle of the main branch.Finally,the growth rate of the Io U of rotated samples is slower than that of horizontal samples,so the cube root calculation is used for skewness to reduce the threshold properly,and the Io U range of calculating threshold is expanded to further reduce the threshold.In this thesis,some Google images are added to the ship data of the DOTA dataset and the average precision AP value on this dataset reaches 58.39%,which is 11.45%better than using TSD directly,and outperforms many other CNN based object detection algorithms.Then we extracted the ship data from the fine grained dataset FAIR1M and modified the label to"ship"to verify the effectiveness of the proposed algorithm,and the final model improved the AP value by 1.40% compared to the RITSD method.To compare with more advanced algorithms,we conducted an experimental comparison on the HRSC2016 dataset,and the model in this thesis achieved 89.9% and 93.1% on the Pascal 2007 and 2012 evaluation metrics,respectively.(5)Based on the framework of ship detection,this thesis adds a ship sub-categorization stage to recognize oil tankers.Besides,the activation function of the gated channel transformation(GCT)is modified to the sigmoid function so that the GCT can be applied in the box head to assign different and greater than 1 weight to highlight the fine-grained features that are more discriminatory.Finally,the average precision AP for liquid cargo ships in the FAIR1M fine-grained dataset reached 59.86%,which is 2.23% better than the AP of the GCT;the average AP for all ships reached 40.29%,which is 1.53% better than the AP of the GCT,and outperformed many other channel attention models.In summary,to predict world oil reserves and the trend of oil reserves,this thesis uses high-resolution optical remote sensing images to estimate the volume of oil tanks and detect oil vessels on the route and port.A calculation method is proposed for the length of the crescent-shaped shadow of oil tanks;and the object detection method based on the convolutional neural network is used to improve the detection accuracy of oil tanks and ships in two aspects:improving the feature region accuracy of classification and regression and dynamic threshold method;finally,the recognition accuracy of oil tankers is improved by using the channel attention mechanism based on the ship detection framework.