Random Forest Variable Importance Research Articles

Data mining approach based on chemical composition of grape skin for quality evaluation and traceability prediction of grapes

The knowledge of wine origin is an important aspect in winemaking industries due to the Denomination of Controlled Origin. In this work, a data mining algorithms comparison study of grape-skin samples from five regions of Mendoza, Argentina, and builds classification models capable of predicting provenance based on multi-elemental composition, were developed. Inductively coupled plasma mass spectrometry (ICP-MS) was used to determine 29 elements (Ag, As, Ce, Co, Cs, Cu, Eu, Fe, Ga, Gd, La, Lu, Mn, Mo, Nb, Nd, Ni, Pr, Rb, Sm, Te, Ti, Tl, Tm, U, V, Y, Zn and Zr). Four classification techniques, including multinomial logistic regression (MLR), k-nearest neighbors (k-NN), support vector machines (SVM), and random forests (RF) were assessed. The best results were achieved for SVM and RF models, with 84% and 88.9% prediction accuracy, respectively, on the 10-fold cross validation. The RF variable importance showed that Rb (rubidium) was the most relevant components for prediction.

Weighted Multi-attribute Framework to Identify Freeway Incident Hot Spots in a Spatiotemporal Context

The present work used the analytical hierarchy process supported by a machine learning technique (namely random forest variable importance) in a geospatial environment to identify traffic-incident hot spots where seven incident attributes (incident severity level, vehicle type, number of vehicles involved, number of weather condition factors, number of lanes blocked, time period, and incident duration) and incident frequency were considered. Using a frame of raster representation, each factor was ranked and weighted based on the literature, expert surveys, and random forest technique to perform hot spot analysis and generate a critical road index map. Statistical analysis of spatial clustering and hot spot spatial densities was carried out based on Moran’s I method of spatial autocorrelation, Getis–Ord Gi* statistics, and point kernel density. Based on weighted attributes of more than 130,000 incidents from Harris County, Texas, for the period between 2004 and 2013, the method successfully identified critical road segments and highlighted factors that contribute to criticality. The results show that high-severity segments are located in the downtown area during the period of 2004 to 2010, and shifted northwest of the downtown area during the period of 2010 to 2013. The random forest variable importance revealed that the number of lanes blocked along with time periods was the most critical factor affecting the severity of traffic incidents. Though the application of the method was demonstrated using incident database from Harris County, it can be generalized to other cities. Traffic agencies can use the suggested approach to garner and visualize reliable information that ensures better traffic management schemes.

Efficacy of Integrating a Novel 16-Gene Biomarker Panel and Intelligence Classifiers for Differential Diagnosis of Rheumatoid Arthritis and Osteoarthritis.

Introducing novel biomarkers for accurately detecting and differentiating rheumatoid arthritis (RA) and osteoarthritis (OA) using clinical samples is essential. In the current study, we searched for a novel data-driven gene signature of synovial tissues to differentiate RA from OA patients. Fifty-three RA, 41 OA, and 25 normal microarray-based transcriptome samples were utilized. The area under the curve random forests (RF) variable importance measurement was applied to seek the most influential differential genes between RA and OA. Five algorithms including RF, k-nearest neighbors (kNN), support vector machines (SVM), naïve-Bayes, and a tree-based method were employed for the classification. We found a 16-gene signature that could effectively differentiate RA from OA, including TMOD1, POP7, SGCA, KLRD1, ALOX5, RAB22A, ANK3, PTPN3, GZMK, CLU, GZMB, FBXL7, TNFRSF4, IL32, MXRA7, and CD8A. The externally validated accuracy of the RF model was 0.96 (sensitivity = 1.00, specificity = 0.90). Likewise, the accuracy of kNN, SVM, naïve-Bayes, and decision tree was 0.96, 0.96, 0.96, and 0.91, respectively. Functional meta-analysis exhibited the differential pathological processes of RA and OA; suggested promising targets for further mechanistic and therapeutic studies. In conclusion, the proposed genetic signature combined with sophisticated classification methods may improve the diagnosis and management of RA patients.

On what to permute in test-based approaches for variable importance measures in Random Forests

In bioinformatics applications, it is currently customary to permute the outcome variable in order to produce inference on covariates to test novel methods or statistics whose distributions are poorly known. The seminal publication of Altmann et al. in Bioinformatics uses the same permutation scheme to obtain P-values that can be treated as corrected measure of feature importance to rectify the bias of the Gini variable importance in Random Forests. Since then, such method has been used in applied work to also draw statistical conclusions on variable importance measures from resulting P-values. In this paper, we show that permuting the outcome may produce unexpected results, including P-values with undesirable properties and illustrate how more refined permutation schemes can be appropriate to obtain desirable results, including high power in discovering relevant variables. Supplementary data are available at Bioinformatics online.

Estimation of forest structural and compositional variables using ALS data and multi-seasonal satellite imagery

Advanced forest resource inventory (FRI) information is of critical importance for sustainable forest management. FRIs are dependent on remote sensing data and processing methods, along with field calibration/validation to generate cost-effective options for modelling forest inventory and biophysical variables over large areas. The objective of this study was to examine the impact of combining multi-seasonal multispectral satellite imagery with airborne laser scanning (ALS) data for estimating basal area, species mixture and stem density for an uneven-aged tolerant hardwood forest in Ontario, Canada. Using random forest (RF) regression as a non-parametric diagnostic technique, three multispectral optical sensors (i.e., Landsat-5 TM, Sentinel-2 A and WorldView-2) were compared to examine the most cost-effective sensor configuration for modelling FRI variables. The contribution of spectral predictors derived from these optical sensors as well as ALS height and intensity metrics were evaluated using RF variable importance. As part of our variable selection framework, all predictor variables were grouped into relatively independent clusters using a hierarchical variable clustering technique, which revealed the distinctiveness between information contained in spectral predictors, height- and intensity-based metrics. This indicates that ALS intensity data carry unique information complementary to passive near-infrared data for forest characterization. ALS data alone did not result in accurate models for basal area and species mixture, but predictive accuracies were improved significantly with the addition of spectral predictors. Compared to single-date images, multi-seasonal imagery proved to be more accurate for modelling FRI variables, especially when combined with ALS data. Despite its limited spatial resolution, Sentinel-2 A was found to be the most cost-effective image source for enhancing ALS-based FRI models. Using variables identified by the variable selection procedure, best subsets regression outperformed the RF models developed for diagnostic analysis, resulting in a suite of accurate and parsimonious predictive models, with coefficients of determination of 0.73, 0.90 and 0.67, for basal area, species mixture, and stem density, respectively.

An efficient feature optimization for wetland mapping by synergistic use of SAR intensity, interferometry, and polarimetry data

Wetlands are home to a great variety of flora and fauna species and provide several unique environmental services. Knowledge of wetland species distribution is critical for sustainable management and resource assessment. In this study, multi-temporal single- and full-polarized RADARSAT-2 and single-polarized TerraSAR-X data were applied to characterize the wetland extent of a test site located in the north east of Newfoundland and Labrador, Canada. The accuracy and information content of wetland maps using remote sensing data depend on several factors, such as the type of data, input features, classification algorithms, and ecological characteristics of wetland classes. Most previous wetland studies examined the efficiency of one or two feature types, including intensity and polarimetry. Fewer investigations have examined the potential of interferometric coherence for wetland mapping. Thus, we evaluated the efficiency of using multiple feature types, including intensity, interferometric coherence, and polarimetric scattering for wetland mapping in multiple classification scenarios. An ensemble classifier, namely Random Forest (RF), and a kernel-based Support Vector Machine (SVM) were also used to determine the effect of the classifier. In all classification scenarios, SVM outperformed RF by 1.5–5%. The classification results demonstrated that the intensity features had a higher accuracy relative to coherence and polarimetric features. However, an inclusion of all feature types improved the classification accuracy for both RF and SVM classifiers. We also optimized the type and number of input features using an integration of RF variable importance and Spearman’s rank-order correlation. The results of this analysis found that, of 81 input features, 22 were the most important uncorrelated features for classification. An overall classification accuracy of 85.4% was achieved by incorporating these 22 important uncorrelated features based on the proposed classification framework.

Poppy crop capsule volume estimation using UAS remote sensing and random forest regression

Improved prediction of poppy capsule volume is essential for optimal management of poppy crop. In order to estimate poppy capsule volume accurately using remotely sensed imagery, the selection of most appropriate models and predictor variables is essential. Multiple spectral indices with random forest (RF) regression were tested to estimate poppy capsule volume using an Unmanned Aircraft System (UAS). Data were collected from field-based physical measurements, in-field spectral measurements and from UAS flights with multispectral sensors over two poppy crops at Cambridge and Sorell in Tasmania, Australia. Field measured spectral signatures were convolved to the multispectral bands of a UAS mounted sensor. These convolved UAS spectral signatures were used to compute multiple spectral indices to develop the RF model, and select optimal model parameters based on root mean squared error (RMSE). In addition, the RF variable importance scores were used to rank the model variables, and to identify the best performing vegetation indices. In Cambridge, an RF model based on convolved UAS spectral signatures predicted capsule volume with an RMSE values ranging from 15.60 cm3 (10.27%) to 25.63 cm3 (14.45%) from training and validation dataset, respectively, indicating a strong relationship between SVIs and field measured capsule volume. An RF model trained on UAS multispectral data (measure not simulated) resulted an RMSE value of 19.39 cm3 (12.80%) based on training data set and an RMSE value of 26.85 cm3 (17.77%) with validation dataset. The Cambridge site model parameters and optimal variables were applied to the Sorell data, which showed a significant relationship between measured and predicted capsule volume (R2 0.72), with relative error of 26.25%. The results showed that the RF model developed using selected variables can help to predict capsule volume 2–3 weeks prior to harvest.

Random Forest Variable Importance Spectral Indices Scheme for Burnt Forest Recovery Monitoring—Multilevel RF-VIMP

Burnt forest recovery is normally monitored with a time-series analysis of satellite data because of its proficiency for large observation areas. Traditional methods, such as linear correlation plotting, have been proven to be effective, as forest recovery naturally increases with time. However, these methods are complicated and time consuming when increasing the number of observed parameters. In this work, we present a random forest variable importance (RF-VIMP) scheme called multilevel RF-VIMP to compare and assess the relationship between 36 spectral indices (parameters) of burnt boreal forest recovery in the Great Xing’an Mountain, China. Six Landsat images were acquired in the same month 0, 1, 4, 14, 16, and 20 years after a fire, and 39,380 fixed-location samples were then extracted to calculate the effectiveness of the 36 parameters. Consequently, the proposed method was applied to find correlations between the forest recovery indices. The experiment showed that the proposed method is suitable for explaining the efficacy of those spectral indices in terms of discrimination and trend analysis, and for showing the satellite data and forest succession dynamics when applied in a time series. The results suggest that the tasseled cap transformation wetness, brightness, and the shortwave infrared bands (both 1 and 2) perform better than other indices for both classification and monitoring.

Early Cerebral Vein After Endovascular Ischemic Stroke Treatment Predicts Symptomatic Reperfusion Hemorrhage.

Parenchymal hemorrhage (PH) after endovascular mechanical thrombectomy in acute ischemic stroke leads to worse outcomes. Better clinical and imaging biomarkers of symptomatic reperfusion PH are needed to identify patients at risk. We identified clinical and imaging predictors of reperfusion PH after endovascular mechanical thrombectomy with attention to early cerebral veins (ECVs) on postreperfusion digital subtraction angiography. We performed a retrospective cohort study of consecutive acute ischemic stroke patients undergoing endovascular mechanical thrombectomy at our neurovascular referral center. Clinical and imaging characteristics were collected from patient health records, and random forest variable importance measures were used to identify predictors of symptomatic PH. Predictors of secondary outcomes, including 90-day mortality, functional dependence (modified Rankin Scale score, >2), and National Institutes of Health Stroke Scale shift, were also determined. Diagnostic test characteristics of ECV for symptomatic PH were determined using a receiver operating characteristic analysis. Differences between patients with and without symptomatic PH were assessed with Fisher exact test and the Wilcoxon rank sum (Mann-Whitney U test) test at the 0.05 significance level. Of 64 patients with anterior circulation large-vessel occlusion identified, 6 (9.4%) developed symptomatic PH. ECV was the strongest predictor of symptomatic PH with more than twice the importance of the next best predictor, male sex. Although ECV was also predictive of 90-day mortality and functional dependence, other characteristics were more important than ECV for these outcomes. The sensitivity and specificity of ECV alone for subsequent hemorrhage were both 0.83, with an area under the curve of 0.83 and 95% confidence interval of 0.66 to 1.00. ECV on postendovascular mechanical thrombectomy digital subtraction angiography is highly diagnostic of subsequent symptomatic reperfusion hemorrhage in this data set. This finding has important implications for post-treatment management of blood pressure and anticoagulation.

MODELLING ABOVE GROUND BIOMASS OF MANGROVE FOREST USING SENTINEL-1 IMAGERY

Abstract. Many studies have been conducted in the estimation of forest above ground biomass (AGB) using features from synthetic aperture radar (SAR). Specifically, L-band ALOS/PALSAR (wavelength ~23 cm) data is often used. However, few studies have been made on the use of shorter wavelengths (e.g., C-band, 3.75 cm to 7.5 cm) for forest mapping especially in tropical forests since higher attenuation is observed for volumetric objects where energy propagated is absorbed. This study aims to model AGB estimates of mangrove forest using information derived from Sentinel-1 C-band SAR data. Combinations of polarisations (VV, VH), its derivatives, grey level co-occurrence matrix (GLCM), and its principal components were used as features for modelling AGB. Five models were tested with varying combinations of features; a) sigma nought polarisations and its derivatives; b) GLCM textures; c) the first five principal components; d) combination of models a−c; and e) the identified important features by Random Forest variable importance algorithm. Random Forest was used as regressor to compute for the AGB estimates to avoid over fitting caused by the introduction of too many features in the model. Model e obtained the highest r2 of 0.79 and an RMSE of 0.44 Mg using only four features, namely, σ°VH GLCM variance, σ°VH GLCM contrast, PC1, and PC2. This study shows that Sentinel-1 C-band SAR data could be used to produce acceptable AGB estimates in mangrove forest to compensate for the unavailability of longer wavelength SAR.

Wisdom of artificial crowds feature selection in untargeted metabolomics: An application to the development of a blood-based diagnostic test for thrombotic myocardial infarction

IntroductionHeart disease remains a leading cause of global mortality. While acute myocardial infarction (colloquially: heart attack), has multiple proximate causes, proximate etiology cannot be determined by a blood-based diagnostic test. We enrolled a suitable patient cohort and conducted a non-targeted quantification of plasma metabolites by mass spectrometry for developing a test that can differentiate between thrombotic MI, non-thrombotic MI, and stable disease. A significant challenge in developing such a diagnostic test is solving the NP-hard problem of feature selection for constructing an optimal statistical classifier. ObjectiveWe employed a Wisdom of Artificial Crowds (WoAC) strategy for solving the feature selection problem and evaluated the accuracy and parsimony of downstream classifiers in comparison with traditional feature selection techniques including the Lasso and selection using Random Forest variable importance criteria. Materials and methodsArtificial Crowd Wisdom was generated via aggregation of the best solutions from independent and diverse genetic algorithm populations that were initialized with bootstrapping and a random subspaces constraint. Results/ConclusionsStrong evidence was observed that a statistical classifier utilizing WoAC feature selection can discriminate between human subjects presenting with thrombotic MI, non-thrombotic MI, and stable Coronary Artery Disease given abundances of selected plasma metabolites. Utilizing the abundances of twenty selected metabolites, a leave-one-out cross-validation estimated misclassification rate of 2.6% was observed. However, the WoAC feature selection strategy did not perform better than the Lasso over the current study.

A Systematic Approach for Variable Selection With Random Forests: Achieving Stable Variable Importance Values

Random Forests variable importance measures are often used to rank variables by their relevance to a classification problem and subsequently reduce the number of model inputs in high-dimensional data sets, thus increasing computational efficiency. However, as a result of the way that training data and predictor variables are randomly selected for use in constructing each tree and splitting each node, it is also well known that if too few trees are generated, variable importance rankings tend to differ between model runs. In this letter, we characterize the effect of the number of trees (ntree) and class separability on the stability of variable importance rankings and develop a systematic approach to define the number of model runs and/or trees required to achieve stability in variable importance measures. Results demonstrate that both a large ntree for a single model run, or averaged values across multiple model runs with fewer trees, are sufficient for achieving stable mean importance values. While the latter is far more computationally efficient, both the methods tend to lead to the same ranking of variables. Moreover, the optimal number of model runs differs depending on the separability of classes. Recommendations are made to users regarding how to determine the number of model runs and/or trees that are required to achieve stable variable importance rankings.

Selecting a subset of spectral bands for mapping invasive alien plants: a case of discriminating Parthenium hysterophorus using field spectroscopy data

ABSTRACTParthenium hysterophorus is considered one of the top seven most problematic and devastating weeds in the world. It compromises the integrity of ecosystems, human health, agricultural production, and biodiversity. Therefore, its early detection and discrimination are critical for facilitating site-specific weed management. Recently, adoption of remote-sensing approaches has gained popularity for species-level mapping of vegetation. Specifically, the use of hyperspectral data has demonstrated reliable mapping accuracy. However, when working with hyperspectral data, feature selection is fundamental to achieving reliable classification accuracies. Moreover, challenges such as ‘the curse of dimensionality’ that cause unstable parameter estimates and high generalization errors when the number of observations (n) is less than the number of descriptive variables (p), i.e. n < p often compromise classification accuracy. In this study, we assessed the potential of a hybrid feature selection approach, based on statistical analysis and Support Vector Machines – Recursive Feature Elimination (SVM-RFE) for determining a subset of hyperspectral bands relevant for discriminating P. hysterophorus using field spectroscopy data. We compared the performance of SVM-RFE, Random Forest variable importance (RF VarImp), and entire spectral dat aset (p = 1633) using SVM classifier with radial basis function (RBF) kernel. Results of SVM-RFE and RF VarImp generated lower classification accuracies (i.e. 76.19% and 66.67%, respectively) than the entire spectral data set, i.e. 78.57%. On the other hand, using a subset of 10 spectral bands, our hybrid approach yielded a superior overall accuracy of 80.19% in discriminating P. hysterophorus from its co-occurring species. The study showed that a subset consisting of two red-edge bands located at 685 and 707 nm, one near infrared band at 1115 nm, and seven short wave infrared bands at 1971, 1982, 1990, 1966, 2003, 2005, and 2013 nm had the greatest potential for discrimination of P. hysterophorus and co-occurring plant groups. Overall, the study suggests that the hybrid approach is effective for early detection and improvement of invasive alien plants classification accuracy, reducing data dimensionality and selecting a relevant spectral subset of bands.

Object-based crop classification using multi-temporal SPOT-5 imagery and textural features with a Random Forest classifier

In this study, an object-based image analysis (OBIA) approach was developed to classify field crops using multi-temporal SPOT-5 images with a random forest (RF) classifier. A wide range of features, including the spectral reflectance, vegetation indices (VIs), textural features based on the grey-level co-occurrence matrix (GLCM) and textural features based on geostatistical semivariogram (GST) were extracted for classification, and their performance was evaluated with the RF variable importance measures. Results showed that the best segmentation quality was achieved using the SPOT image acquired in September, with a scale parameter of 40. The spectral reflectance and the GST had a stronger contribution to crop classification than the VIs and GLCM textures. A subset of 60 features was selected using the RF-based feature selection (FS) method, and in this subset, the near-infrared reflectance and the image acquired in August (jointing and heading stages) were found to be the best for crop classification.

A Fully Polarimetric SAR Imagery Classification Scheme for Mud and Sand Flats in Intertidal Zones

Sediments on exposed intertidal flats are very dynamic and perform vital ecosystem functions. This paper proposes a new classification scheme for mud and sand flats on intertidal flats using fully polarimetric synthetic aperture radar (SAR) data. Freeman–Durden (FD) and Cloude–Pottier (CP) polarimetric decomposition components as well as double bounce eigenvalue relative difference (DERD) are introduced into the feature sets instead of the original intensity polarimetric channels. Classification is carried out using the random forest (RF) theory, and the results are evaluated using confusion matrices, kappa coefficients, and RF variable importance indices. Three study sites with different environmental conditions are chosen to demonstrate the effectiveness of the proposed classification chain. To further assess the performance of the proposed feature set, we set different feature combinations and process with the same processing chain. Results show that the DERD parameter can detail the sediment mappings on exposed intertidal flats and is a useful SAR feature to distinguish mud and sand flats efficiently. The combined FD and CP components have the ability to describe the polarimetric characteristics of sediments more correctly than the commonly used original intensity channels. Meanwhile, the RF theory shows great potential in distinguishing sediments in intertidal zones accurately and time efficiently.

Using conditional inference forest to identify variable importance

Variable importance measure with Random Forests (RF) have received increased attention as a means of variable selection in classification tasks. The measure of variable importance in Random Forests is a smart way of variable selection in many applications, but is not reliable in situations where potential predictor variables vary in their scale of measurement or their number of categories. In this paper, we have implemented Random Forest built from Conditional Inference Trees (CIT) that is called Conditional Inference Forest (CIF). In each tree in the forest of conditional inference, the division of the nodes is based on the way to have a good associativity. The chi-square test statistics is used to measure the association. In addition to identifying variables that improve the classification accuracy, the methodology also clearly identifies the variables that are neutral to the accuracy, and also those who interfere in the right classification. In this paper, we are particularly interested in the overall algorithm Conditional Inference Forest (CIF) for the classification of large biological data. The algorithm is evaluated on its ability to select a reduced number of features while preserving a very satisfactory classification rate.

Factors Associated with HIV Testing Among Participants from Substance Use Disorder Treatment Programs in the US: A Machine Learning Approach.

HIV testing is the foundation for consolidated HIV treatment and prevention. In this study, we aim to discover the most relevant variables for predicting HIV testing uptake among substance users in substance use disorder treatment programs by applying random forest (RF), a robust multivariate statistical learning method. We also provide a descriptive introduction to this method for those who are unfamiliar with it. We used data from the National Institute on Drug Abuse Clinical Trials Network HIV testing and counseling study (CTN-0032). A total of 1281 HIV-negative or status unknown participants from 12 US community-based substance use disorder treatment programs were included and were randomized into three HIV testing and counseling treatment groups. The a priori primary outcome was self-reported receipt of HIV test results. Classification accuracy of RF was compared to logistic regression, a standard statistical approach for binary outcomes. Variable importance measures for the RF model were used to select the most relevant variables. RF based models produced much higher classification accuracy than those based on logistic regression. Treatment group is the most important predictor among all covariates, with a variable importance index of 12.9%. RF variable importance revealed that several types of condomless sex behaviors, condom use self-efficacy and attitudes towards condom use, and level of depression are the most important predictors of receipt of HIV testing results. There is a non-linear negative relationship between count of condomless sex acts and the receipt of HIV testing. In conclusion, RF seems promising in discovering important factors related to HIV testing uptake among large numbers of predictors and should be encouraged in future HIV prevention and treatment research and intervention program evaluations.

Enhancing Multimedia Imbalanced Concept Detection Using VIMP in Random Forests.

Recent developments in social media and cloud storage lead to an exponential growth in the amount of multimedia data, which increases the complexity of managing, storing, indexing, and retrieving information from such big data. Many current content-based concept detection approaches lag from successfully bridging the semantic gap. To solve this problem, a multi-stage random forest framework is proposed to generate predictor variables based on multivariate regressions using variable importance (VIMP). By fine tuning the forests and significantly reducing the predictor variables, the concept detection scores are evaluated when the concept of interest is rare and imbalanced, i.e., having little collaboration with other high level concepts. Using classical multivariate statistics, estimating the value of one coordinate using other coordinates standardizes the covariates and it depends upon the variance of the correlations instead of the mean. Thus, conditional dependence on the data being normally distributed is eliminated. Experimental results demonstrate that the proposed framework outperforms those approaches in the comparison in terms of the Mean Average Precision (MAP) values.

RANDOM FOREST CLASSIFICATION OF SEDIMENTS ON EXPOSED INTERTIDAL FLATS USING ALOS-2 QUAD-POLARIMETRIC SAR DATA

Coastal zones are one of the world’s most densely populated areas and it is necessary to propose an accurate, cost effective, frequent, and synoptic method of monitoring these complex ecosystems. However, misclassification of sediments on exposed intertidal flats restricts the development of coastal zones surveillance. With the advent of SAR (Synthetic Aperture Radar) satellites, polarimetric SAR satellite imagery plays an increasingly important role in monitoring changes in coastal wetland. This research investigated the necessity of combining SAR polarimetric features with optical data, and their contribution in accurately sediment classification. Three experimental groups were set to make assessment of the most appropriate descriptors. (i) Several SAR polarimetric descriptors were extracted from scattering matrix using Cloude-Pottier, Freeman-Durden and Yamaguchi methods; (ii) Optical remote sensing (RS) data with R, G and B channels formed the second feature combinations; (iii) The chosen SAR and optical RS indicators were both added into classifier. Classification was carried out using Random Forest (RF) classifiers and a general result mapping of intertidal flats was generated. Experiments were implemented using ALOS-2 L-band satellite imagery and GF-1 optical multi-spectral data acquired in the same period. The weights of descriptors were evaluated by VI (RF Variable Importance). Results suggested that optical data source has few advantages on sediment classification, and even reduce the effect of SAR indicators. Polarimetric SAR feature sets show great potentials in intertidal flats classification and are promising in classifying mud flats, sand flats, bare farmland and tidal water.

RANDOM FOREST CLASSIFICATION OF SEDIMENTS ON EXPOSED INTERTIDAL FLATS USING ALOS-2 QUAD-POLARIMETRIC SAR DATA

Abstract. Coastal zones are one of the world’s most densely populated areas and it is necessary to propose an accurate, cost effective, frequent, and synoptic method of monitoring these complex ecosystems. However, misclassification of sediments on exposed intertidal flats restricts the development of coastal zones surveillance. With the advent of SAR (Synthetic Aperture Radar) satellites, polarimetric SAR satellite imagery plays an increasingly important role in monitoring changes in coastal wetland. This research investigated the necessity of combining SAR polarimetric features with optical data, and their contribution in accurately sediment classification. Three experimental groups were set to make assessment of the most appropriate descriptors. (i) Several SAR polarimetric descriptors were extracted from scattering matrix using Cloude-Pottier, Freeman-Durden and Yamaguchi methods; (ii) Optical remote sensing (RS) data with R, G and B channels formed the second feature combinations; (iii) The chosen SAR and optical RS indicators were both added into classifier. Classification was carried out using Random Forest (RF) classifiers and a general result mapping of intertidal flats was generated. Experiments were implemented using ALOS-2 L-band satellite imagery and GF-1 optical multi-spectral data acquired in the same period. The weights of descriptors were evaluated by VI (RF Variable Importance). Results suggested that optical data source has few advantages on sediment classification, and even reduce the effect of SAR indicators. Polarimetric SAR feature sets show great potentials in intertidal flats classification and are promising in classifying mud flats, sand flats, bare farmland and tidal water.