Search

China has been facing serious land degradation and desertification in its north and northwest arid and semi-arid areas. Monitoring the dynamics of percentage vegetation cover (PVC) using remote sensing imagery in these areas has become critical. However, because these areas are large, remote, and sparsely populated, and also because of the existence of mixed pixels, there have been no accurate and cost-effective methods available for this purpose. Spectral unmixing methods are a good alternative as they do not need field data and are low cost. However, traditional linear spectral unmixing (LSU) methods lack the ability to capture the characteristics of spectral reflectance and scattering from endmembers and their interactions within mixed pixels. Moreover, existing nonlinear spectral unmixing methods, such as random forest (RF) and radial basis function neural network (RBFNN), are often costly because they require field measurements of PVC from a large number of training samples. In this study, a cost-effective approach to mapping PVC in arid and semi-arid areas was proposed. A method for selection and purification of endmembers mainly based on Landsat imagery was first presented. A probability-based spectral unmixing analysis (PBSUA) and a probability-based optimized k nearest-neighbors (PBOkNN) approach were then developed to improve the mapping of PVC in Duolun County in Inner Mongolia, China, using Landsat 8 images and field data from 920 sample plots. The proposed PBSUA and PBOkNN methods were further validated in terms of accuracy and cost-effectiveness by comparison with two LSU methods, with and without purification of endmembers, and two nonlinear approaches, RF and RBFNN. The cost-effectiveness was defined as the reciprocal of cost timing relative root mean square error (RRMSE). The results showed that (1) Probability-based spectral unmixing analysis (PBSUA) was most cost-effective and increased the cost-effectiveness by 29.3% 29.3%, 33.5%, 50.8%, and 53.0% compared with two LSU methods, PBOkNN, RF, and RBFNN, respectively; (2) PBSUA, RF, and RBFNN gave RRMSE values of 22.9%, 21.8%, and 22.8%, respectively, which were not significantly different from each other at the significance level of 0.05. Compatibly, PBOkNN and LSU methods with and without purification of endmembers resulted in significantly greater RRMSE values of 27.5%, 32.4%, and 43.3%, respectively; (3) the average estimates of the sample plots and predicted maps from PBSUA, PBOkNN, RF, and RBFNN fell in the confidence interval of the test plot data, but those from two LSU methods did not, although the LSU with purification of endmembers improved the PVC estimation accuracy by 25.2% compared with the LSU without purification of endmembers. Thus, this study indicated that the proposed PBSUA had great potential for cost-effectively mapping PVC in arid and semi-arid areas. [ABSTRACT FROM AUTHOR]

Traditional parametric methods for classification of land use and land cover (LULC) types using remote sensing imagery assume a global distribution model and fail to consider local variation of categorical variables. Differently, non-parametric methods do not make any statistical assumptions but are typically sensitive to the sample sizes of training sample data that usually require a high cost to collect in the field. Geostatistical classifiers, such as indicator kriging and simulation, are local variability-based methods that exhibit great potential for image-based classification of LULC types. However, variogram models required are highly sensitive to the spatial configuration of training samples as well as sample size given a study area. Moreover, when a large number of spectral variables are considered into kriging systems, modeling the variograms and cross-variograms would be problematic. To circumvent these issues, this study extended the geostatistical methods from a 2-dimensional geographic space to a m -dimensional image feature space to derive feature-space indicator variograms (FSIVs). Moreover, a novel stochastic simulation classification algorithm, Feature-Space Indicator Simulation (FSIS), was proposed and examined for classification of LULC types in Duolun County located in Inner Mongolia and in Huang-Feng-Qiao (HFQ) forest farm, Hunan of China. In Duolun, six LULC types were involved and in HFQ a complicated forest landscape consisting of nine forest types plus water, built-up area, and agricultural/bare soil, was classified. The classification results of FSIS were compared with another feature-space geostatistical classifier – feature-space indicator kriging (FSIK), a traditional parametric method – maximum likelihood (ML), a widely used nonparametric method – support vector machine (SVM), and a recently popular method – random forest (RF). The results showed that compared with ML, SVM and RF, in both study areas FSIS statistically significantly increased the accuracy of the classifications by 10.0–29.9% for percentage correct and 19.0–47.6% for Kappa statistic. Compared with FSIK, FSIS also improved the classification accuracy but the accuracy increases were relatively smaller with the percentages correct of 3.5% and 7.6% and the Kappa values of 4.6% and 8.6% for Duolun and HFQ, respectively. Moreover, FSIS led to the spatial uncertainties of the classification estimates as the quality measure of the estimates. In addition, the results also demonstrated that FSIVs were sensitive to the within-class heterogeneity but not very much to the size of training samples. Overall, FSIS exhibited the greater potential to improve the classification accuracy of LULC and forest types using remote sensing image. [ABSTRACT FROM AUTHOR]

Forest growing stem volume (GSV) reflects the richness of forest resources as well as the quality of forest ecosystems. Remote sensing technology enables robust and efficient GSV estimation as it greatly reduces the survey time and cost while facilitating periodic monitoring. Given its red edge bands and a short revisit time period, Sentinel-2 images were selected for the GSV estimation in Wangyedian forest farm, Inner Mongolia, China. The variable combination was shown to significantly affect the accuracy of the estimation model. After extracting spectral variables, texture features, and topographic factors, a stepwise random forest (SRF) method was proposed to select variable combinations and establish random forest regressions (RFR) for GSV estimation. The linear stepwise regression (LSR), Boruta, Variable Selection Using Random Forests (VSURF), and random forest (RF) methods were then used as references for comparison with the proposed SRF for selection of predictors and GSV estimation. Combined with the observed GSV data and the Sentinel-2 images, the distributions of GSV were generated by the RFR models with the variable combinations determined by the LSR, RF, Boruta, VSURF, and SRF. The results show that the texture features of Sentinel-2's red edge bands can significantly improve the accuracy of GSV estimation. The SRF method can effectively select the optimal variable combination, and the SRF-based model results in the highest estimation accuracy with the decreases of relative root mean square error by 16.4%, 14.4%, 16.3%, and 10.6% compared with those from the LSR-, RF-, Boruta-, and VSURF-based models, respectively. The GSV distribution generated by the SRF-based model matched that of the field observations well. The results of this study are expected to provide a reference for GSV estimation of coniferous plantations. [ABSTRACT FROM AUTHOR]