| Accurate identification of rock types provides important research materials for the study of geological evolution history,resource exploration,geotechnical engineering,and other related fields.According to different rock formation processes,the composition,texture,color,and other characteristics are often different,and these differences are important in distinguishing rocks.Limited by surface conditions and other factors,traditional field investigation is time-consuming and costly,while remote sensing technology can rapidly observe the surface on a large scale without being affected by surface conditions,which greatly improves the efficiency of geology interpretation,provides valid research data for rock classification,and promotes the process of high-precision and quantitative rock interpretation.The development of computer science,statistics,and mathematical science has promoted rock classification to develop from manual visual interpretation to semi-automatic and automatic classification.However,the limitation of sensor performance and the complexity of geological bodies still restrict the improvement of rock classification accuracy.At present,rock classification using remote sensing data still faces a series of problems.For instance,the influence of external factors such as illumination and weather on the spectral characteristics of rocks cannot be effectively eliminated,and the large amount of speckle noise in radar images obscures the difference between radar features of different rocks,and the difficulty in carrying out rock classification research in vegetated areas,etc.Under the current technical level,how to improve the rock classification accuracy of single-source remote sensing data needs further exploration,using the multi-source remote sensing data to explore more image features that can be used for rock classification is of great significance.This study aims at exploring rock image features and improving rock classification accuracy with single-source remote sensing data.Automatic rock classification was carried out in the rock outcrop region and the vegetated region,respectively.Considering the environmental factors,we use multi-temporal optical and radar images(Sentinel-1,Landsat-8,and Sentinel-2)to extract different features,and then combined these features with different classification algorithms for rock classification.The main achievements are as follows:1.We explore the potential of Sentinel-1 dual-polarization radar data to distinguish different rocks,extract the backscatter coefficient,texture,and polarization decomposition features of rocks,and use a variety of classification algorithms such as partial least squares discriminant analysis,support vector machine algorithm,etc.to distinguish different rocks(andesite,dolomite,granite,sandstone,and limestone),and the performance of different classifiers is compared under the condition of small sample size.The results show that:(1)dual-polarization radar data has a good discriminative ability for specific rock types(such as limestone);(2)radar texture features show great potential for rock classification;(3)there is no significant difference between the backscatter coefficient and polarization decomposition features extracted from single-date radar data among different rocks;(4)for small sample sets,the partial least squares discriminant analysis has the highest precision.2.We propose to classify rock units using multi-temporal Sentinel-1 data and random forest classifiers.The backscatter coefficient and the backscattering mechanism of rocks under different incidence angles are compared,and the appropriate incidence angles are selected for subsequent classification.Extract the multi-temporal backscatter coefficient and coherence features of rock units,conduct statistical and temporal analysis of radar features,and then combine the radar features with random forest classifiers to classify rock units.Analyze the influence of window size,direction,and distance of texture features on rock unit classification.The results show that:(1)using radar data with a large incident angle can enhance the discrimination between rock units;(2)the multi-temporal backscatter coefficients and coherence features among different rock units have significant differences;(3)compare with single-date radar data,multi-temporal radar data can improve the classification accuracy by about 10% and can demonstrate the influence of rainfall on the radar features of rock units;(4)for rock unit classification,the backscatter coefficient and coherence feature can complement each other,and different polarizations can also complement each other;(5)appropriate window size,direction,and distance can make texture features achieve a high classification accuracy by reducing the interference of speckle noise.3.We propose to classify rock units using multi-temporal Landsat-8 data and random forest classifiers.Extract multi-temporal reflectance,temperature(LST),brightness(TCTB),greenness(TCTG),and wetness(TCTW)features,analyze the temporal variation of features,and then use different features and feature combinations to classify rock units.The research results show that:(1)the multi-temporal LST,TCTG,TCTB,and TCTW features extracted from Landsat-8 data have significant differences among different rock units;(2)compare with the reflectance from single-date,multi-temporal reflectance can obtain higher classification accuracy;(3)multitemporal TCTB+TCTG+TCTW can obtain comparable or even better classification results than multi-temporal reflectance;(4)when only considering one single multi-temporal feature,the classification accuracy of the LST feature is the highest;(5)multi-temporal LST+TCTB+TCTG+TCTW achieves the highest accuracy(85.26%)with minimum noise content.4.We propose to use time-series Sentinel-2 data with the random forest classifiers and Gradient Boosting classifiers to classify the underlying bedrock in the vegetated area.Analyze and compare the spatiotemporal change of spectral features of vegetation on different bedrocks,and then combine vegetation features with classifiers to classify bedrocks.The results show that:(1)the optical features of vegetation on different bedrocks behave differently during the growth period,and the differences between the optical features of vegetation can be observed;(2)during the mature period,the vegetation on volcanic rocks grows best,and vegetation on terrigenous clastic rocks grows worst;(3)when the growth period transit to the dormant period,vegetation on volcanic rocks enter the dormancy period fastest,and when the dormancy period transit to the growth period,the vegetation on carbonate rocks and intermediate-acid intrusive rocks enter the growth period fastest(4)during the vegetation growth period,the classification result obtained by the vegetation reflectance is better than using vegetation indices;(5)the bedrock classification accuracy of the Gradient Boosting classifiers is better than that of the random forest classifiers(the accuracy is increased by about 4%);(6)the highest bedrock classification accuracy(77.66%)is obtained by the combined use of vegetation reflectance in the growing season and the Gradient Boosting classifier.The bedrock classification results obtained by using the time-series optical features of vegetation are consistent with the reference geological map,which provides a reliable basis for geological interpretation and geological mapping of high vegetation coverage areas. |
PDF Full Text Request