Research On Detection Of Frozen Tomb Targets In High Resolution Remote Sensing Images Based On Deep Learning

The Altai Mountains are one of the most impressive and valuable archaeological areas in the world.The frozen tombs of ancient civilizations,which scatter across the whole Altai area,are exceptionally valuable source of information for archaeology study.These precious archaeological resources,which have been preserved intact in the permafrost underground for over two millenniums,are now under various threats,such as natural disasters,farmland expansion,touristic development,and most notably the global warming.A detailed map or inventory of the frozen tombs is actually essential,but is still not available at now.In the past,archaeologists always used blind field surveys to find frozen tombs,which took time and effort.Archaeological investigations have only been carried out on a small scale of Altai Mountains.Therefore,how to quickly and accurately find frozen tombs is an urgent task.This is a difficult first step in the archeological work of frozen tombs.Deep learning target detection has been widely used in various scenarios due to its excellent effect and efficiency,but it is the first attempt in frozen tomb archaeology with deep learning.High-resolution remote sensing satellite images cover a wide range,and the frozen tombs can exhibit certain characteristics,making it possible to quickly detect frozen tombs.However,there is no suitable frozen tomb detection network model so far.At the same time,the size of frozen tombs in the remote sensing images are extremely small,the background is complex and the contrast with the background is also very low,which greatly increases the difficulty of detection.In this paper,we validate the deep convolutional neural network(CNN)technique for automatic detection of frozen tombs from high resolution satellite images in four Altai areas.Our results demonstrate that CNN is feasible to be applied in frozen tombs detection,but some specific improvement techniques are needed to increase the performance.On the one hand,the near infrared band is added to the data set to increase the characteristics of the frozen tombs in different geomorphologies.On the other hand,a typical complex background training data set is made to increase the recognition degree of the frozen tombs and the background.Due to the extreme imbalance between the positive and negative training samples during actual training,the training bias is serious.We solve it from two directions: First,reduce the step size of anchor to increase the generation of candidate boxes,which can get more and better positive samples.Second,excessive negative samples are dynamically adjusted according to the number of positive samples,which greatly reduces the number of negative samples.Although the recall and precision are still not very high,we observed that the detection results obtained by the proposed method are already to sufficiently capture the spatial distribution of the frozen tombs.It can largely narrow the search areas for the archaeologists in unknown regions,and be useful in preparing field survey campaigns and directing archaeological fieldworks.We also applied the proposed method in an unknown Altai Mountain area and successfully discovered some new frozen tombs there,which are undocumented to the archaeology society.Besides,the proposed method has potential to be applied to construct an inventory for all the scattered frozen tombs in the whole Altai Mountains.