Extracting Ground Fissures In Loess Landform Area Using Modified MF-FDOG Algorithm And UAV Images

High-speed railway,whose operational safety is the focus that people concern about in a long term,has been generally become main vehicle of passenger transportation in China.However,the topography of the Loess Plateau is very complex and varied,geologic hazards(e.g.,landslide and collapse)caused by unstable loess have a great threat to high-speed traffic network and their subsidiary facilities(e.g.,tunnel,culvert and bridge).On the order hand,Loess Plateau is an important base of energy in our country,and mineral resources such as petroleum,coal,rare metals are widely distributed.Accidents caused by mining mineral resources are numerous.In view of this,it is critical important to monitor geologic hazards for operational safety of traffic transportation and mining area.Ground fissures are common geologic phenomenon,which is often accompanied by landslide and surface subsidence.Unmanned aerial vehicle(UAV)has characteristics of strong flexibility,manipulation and low cost.It can receive image with ultra-high spatial resolution.Therefore,the landslide and the collapse can be dynamically monitored by using matched filter with first-order derivative of the Gaussian(MF-FDOG)algorithm to extract ground fissures from UAV images in different periods of time.This method,however,has a disadvantage that the generality of sensitivity correction parameter is poor when modifying the range between two response images came from matched filter(MF)algorithm and firstorder derivative of the Gaussian(FDOG)algorithm,respectively.In addition,incorrect ground fissure targets resulted from withered grass and bare loess etc were not able to remove in the MF-FDOG algorithm.It is indicated that this algorithm is not adaptive for ground fissure extraction of UAV image in the Loess Plateau in China.According to the disadvantage of sensitivity correction parameter in exiting MF-FDOG algorithm,we propose a modified MF-FDOG algorithm.Firstly,creating a group of MF templates and a group of FDOG templates that satisfy vertical profile curve of ground fissure in terms of their geometric attribution in UAV images and then respectively filtering UAV image to enhance signal strength of ground fissure with these two group of templates.Then using linear stretching to eliminate the large differences of range between MF response image and FDOG response image came from filtering operation,and making the difference of two response images after linear stretching to enhance the signals of ground fissure and weaken the signals of other object edges.According to statistics,the histogram of grayscale in differential image satisfies Gaussian distribution,is used as threshold of segmentation to extract ground fissure in UAV images.In the light of simulation experiment of ground fissure,the results showed that the algorithm we proposed has a strong generality compared with using sensitivity correction parameter.From the research of the number of template direction,we can conclude that the best number of template direction in ground fissure extraction is 10.In order to remove incorrect targets of ground fissure extraction,random forest(RF)algorithm is used to classify image and then convert vegetation and bare loess etc category into mask files to delete incorrect targets.According to experiment of classification in UAV image,83% overall accuracy and 77.36% Kappa coefficient could be achieved,respectively.The results are highly consistent with referenced data.For improving accuracy of ground fissures extraction,hit-or-miss transform algorithm is adopted to connect broken ground fissures.At the same time,a filter algorithm we propsed is used to improve the final result of ground fissures by combined with shape and area of small pixel patch to.An experiment was conducted to prove the effect.This paper uses UAV images of mountain around a tunnel in Haidong city,Qinghai province and Yaojie minor district in Lanzhou city,Gansu province as research data and uses modified MF-FDOG algorithm to extract ground fissure.Then comparing with the results of Canny edge detection algorithm,grayscale threshold segmentation algorithm and support vector machine(SVM)algorithm,respectively.In the last,receiver operating characteristic curve(ROC)and overall accuracy are used for accuracy assessment.The results of the experiment showed that the modified MF-FDOG algorithm has a strong applicability of extraction to UAV images acquired in different periods and regions.Compared with common algorithms about linear target extraction,this algorithm achieves a prior accuracy of ground fissure extraction.Therefore,the algorithm we proposed provides an effective technology support to monitor and prevent geological hazards(e.g.,landslides).