Intelligent Identification Of Typical Geo-environmental Information Using New Remote Sensing Data

There are a wide range of geo-environmental problems (GEPs) affecting extensive areas with high frequency across the world, which are detrimental to human survival and sustainable development. Therefore, the identification of earth surface information presented in geo-environments (hereafter referred to as earth surface information) considered as one of the geo-environmental information is of great significance to geo-environment protection and ecological civilization construction. The dissertation focused on the intelligent identification of several typical geo-environmental information (TGEI) which was defined hereby as the earth surface information aforementioned:(1) forested landslide detection in the Three Gorges region characterized by bedrock-covered mountains with steep and rugged terrain, (2) land cover mapping (LCM) in complex surface-mined and agricultural landscapes (CSMAL) located in Wuhan City of central China, (3) land cover classification (LCC) in groundwater dependent ecosystems (GDEs) of an inland arid region located in Dunhuang Basin of northwestern China.Remote sensing techniques have been increasingly playing a vital role for the identification of earth surface information. However, using traditional remote sensing data such as Landsat to identify complex earth surface information have some limitations due to the complex geological setting, geological process, and drastic human disturbance. Recently, several new remote sensing data (NRSD) such as light detection and ranging (LiDAR) (providing high accuracy and bare earth topographic data) and stereo satellite sensor data (providing both high spatial resolution optical image and topographic data) were developed and widely employed. They could potentially improve the identification capability of complex earth surface information.First, three NRSD, i.e., LiDAR, ZiYuan-3 (ZY-3), and RapidEye, were selected. Then to achieve the intelligent identification of the three TGEI, an effective procedure involving the following key points was developed and applied:(1) developing more effective features from the utilized remote sensing data; (2) performing feature reduction (FR) methods to obtain effective feature subsets, involving two wrapper feature selection (FS) methods (i.e., varSelRF and Boruta packages) and three feature extraction (FE) methods (i.e., principal component analysis, PCA; independent component analysis, ICA; minimum noise fraction, MNF); (3) employing three machine learning algorithms (MLAs) (i.e., random forest, RF; support vector machine, SVM; artificial neural network, ANN) with great performance for high accuracy classification, and (4) using semi-automated methods for identifying landslide boundaries based on pixel-based image analysis (PBIA) and object-based image analysis (OBIA).To our knowledge, there were two key scientific issues in the dissertation:(1) how to build up effective features for the characterization of the TGEI, (2) how to pick out the effective feature subsets from the high dimension feature sets for identification of the TGEI. The dissertation developed some novel aspects accordingly.(1) A number of texture and filter features of topographic variables and filter features of spectral bands were developed based on the high resolution topographic and optical data. They can be used as sensitive features and combination to improve the classification accuracies of forested landslides in bedrock-covered mountains with steep and rugged terrain and land cover information in CSMAL.(2) A feature subset construction method based on varSelRF and a selected time threshold of features was proposed. It used multiple FS procedures and the selected time threshold to solve the issue that the selected feature subsets may vary with different training data. It could contribute to get better classification accuracies of forested landslides in bedrock-covered mountains with steep and rugged terrain and land cover information in CSMAL and meanwhile reduce the dimension of data to improve the computing efficiency.The main tasks of the dissertation were as follows.1. Forested landslide detection in bedrock-covered mountains with steep and rugged terrainIt was the first study for forested landslide detection in bedrock-covered mountains with steep and rugged terrain with only LiDAR derivatives. PBIA and OBIA were investigated.And the following new findings were drawn. (1) The new pixel features such as the aspect, digital terrain model (DTM), and slope textures based on aspect direction and the object layer features of Max and Min, and newly introduced object features calculated from the filter features of surface roughness were developed and revealed to improve the classification accuracy. (2) The effect of FS on classification accuracy and the comparisons of OBIA and PBIA, two MLAs and their respective sensitivities to FS, were investigated:(a) FS improved classification accuracy and reduced features for both PBIA and OBIA; (b) RF algorithm achieved higher accuracy and was less sensitive to FS than SVM; (c) compared to PBIA, OBIA was more sensitive to FS, further reduced computing time, and depicted more contiguous terrain segments. (3) The intermittent landslide boundaries delineated using the Canny operator were consistent with the referenced landslide inventory maps; the landslide boundaries derived by the newly developed semi-automatic delineation method were consistent with the referenced landslide inventory maps, with a position mismatch value of 9%.In general, forested landslides in the steep and rugged terrain were intelligently identified by using LiDAR data and the outlined effective procedures based on PBIA and OBIA.2. LCM in CSMALIt was the first study for LCM in CSMAL, involving two tasks:LCM of surface-mined and agricultural landscapes (MSMAL) and classification of surface-mined land (CSML).The following new findings were drawn. (1) The effectiveness of features:(a) textures were of little use, and the novel features e.g., the mean and standard deviation filters of spectral bands and the used topographic features derived from ZY-3 stereo images helped to improve classification accuracy; (b) DTM and normalized difference vegetation index (NDVI) achieved higher importance. (2) The effect of FS on the accuracies of MSMAL and CSML, the comparison of three MLAs’ respective sensitivities to FS, and whether FS resulted in significant influences, were examined:(a) FS substantially reduced feature set and often improved the overall classification accuracies (with an average value of 4.48% for MSMAL using three MLAs, and 11.39% for CSML using RF and SVM); (b) FS showed statistically significant improvements except for ANN-based models for MSMAL; (c) SVM was most sensitive to FS, followed by ANN, and RF; (d) land covers classified easier (e.g. opencast stope and road) were less sensitive to FS. (3) Comparison of three MLAs and whether there were statistically significant differences among them for MSMAL and CSML were examined:(a) for MSMAL based on feature subset, RF achieved greatest overall accuracy (OA) of 77.57%, followed by SVM, and ANN; for CSML, SVM had highest accuracies (87.34%), followed by RF, and ANN; (b) based on the feature subsets, significant differences were observed for MSMAL and CSML using any pair of MLAs.In general, LCM maps in CSMAL were achieved by using ZY-3 imagery and the proposed procedure.3. LCC in GDEs of an inland arid regionThis study focused on following three aspects:LCC based on RapidEye image’s five bands and two VIs i.e., NDVI and its red-edge adaptation (NDVI_RE) by using RF algorithm; assessing the effects of red-edge band and VIs (NDVI and NDVI_RE) on LCC accuracy using feature sets including or excluding them; comparison and integration of FS and FE methods for improving LCC accuracy.The new findings were drawn as follows. (1) High accuracy LCC map (with an OA of 89.67%) achieved based on RapidEye imagery and RF algorithm. (2) The effects of red-edge band and VIs (NDVI and NDVI_RE1) on LCC accuracy:(a) the red-edge band only slightly increased the LCC accuracy; (b) VIs showed significant positive effect on LCC accuracy; (c) simultaneous addition of the red-edge band and VIs achieved a significant OA improvement (3.46% beginning with 86.67%). (3) Comparison and integration of FS and FE methods for improving LCC accuracy:(a) the novel red-edge simple ratio and green NDVI could provide effective information for LCC; (b) all the five FR methods could improve classification accuracy, but only varSelRF achieved significant improvement; (c) varSelRF outperformed the FE methods, followed by MNF, PCA, and ICA; (d) the newly developed integrated varSelRF-PCA model significantly improved classification accuracy (2.66% beginning with 88.17%) and outperformed all the FS or FE methods.In general, using RapidEye imagery and the proposed procedure high accuracy LCC were achieved and the developed red-edge VIs and integrated FR method helped to improve the classification accuracy.In conclusion, the TGEI could be intelligently and effectively identified using NRSD and the procedure proposed in this dissertation. The results drew in the present study will contribute to geo-environment protection and ecological civilization construction.