support vector machine

The aim of this study is to investigate the performance of various algorithms in enhancing the accuracy of land cover maps using Sentinel images.

The study area is a sheet (1:25000) encompassing an area of 15782.6 hectares. The land cover map was created using a range of algorithms - Mahalanobis distance, maximum likelihood, minimum distance, neural network, parallelepiped, support vector machine, spectral angle mapper, spectral information divergence, and binary encoding - applied to the optimal band composition derived from 12 bands (2, 3, 4, 5, 6, 7, 8, 8a, 11, 12, NDVI, SAVI). The training area was obtained from field information and satellite images. 70% of the samples were used for classification and 30% for evaluating the accuracy of the classified maps.

The results indicate that the support vector machine, neural network, and maximum likelihood classifications have the highest accuracy. The support vector machine classifier is slightly more accurate than the other two classifiers. To improve the classification, support vector machine kernels (linear, polynomial, radial basis function, and sigmoid), spectral angle mapper settings (6 modes), and parallelepiped (2 modes) were utilized. The results show that the support vector machine classifier, using the 6th degree polynomial function method, has the highest accuracy. Following this, the map was post-processed with the highest accuracy, resulting in an overall accuracy of 96.63 and a kappa coefficient of 0.9393.

An examination of 21 classification modes on Sentinel images in this study provides a comprehensive review of classification algorithms compared to previous studies.

Effective mangrove mapping needs reliable indices that can characterize mangroves from other land cover types on remote sensing data. Currently, a variety of vegetation- and mangrove indices are available, while a comparative assessment of their efficiency seems essential. The aim of this study was to evaluate six vegetation indices and seven mangrove indices developed for Sentinel-2 imagery to obtain a robust approach in mangrove mapping within Google Earth Engine (GEE) cloud computing platform. The rasterized indices were classified by support vector machine. The final maps were evaluated by area under curve (AUC) of receiver operating characteristic (ROC) in addition to common accuracy assessment criteria. Results showed that mangrove indices were more reliable than vegetation indices. Amongst the vegetation indices, Modified Chlorophyll Absorption in Reflectance Index (MCARI) (AUCmangrove from 0.91 to 0.92) achieved the highest AUC values, while MFI (Mangrove Forest Index) returned the highest values amongst the mangrove indices (AUCmangrove from 0.93 to 0.95). All in all, results revealed that MFI on Sentinel-2 imagery in GEE was efficient in mangrove mapping within the study sites.

Information on forest types and their spatial distribution are valuable for sustainable forest management and planning. The use of remote sensing technology and geographic information system for providing such fundamental information specially in mountainous and remote areas, has been considered by many researchers and forest managers. The current study aims to investigate the capability of Landsat-8 and Sentinel-2 satellite data to generate forest type map in the Kojur watershed of Hyrcanian forests. The performance of some parametric and non-parametric classification methods was also compared.

Following quality assessment, some preprocessing techniques including vegetation indices (VI) extraction, tasseled cap transformation (TCT), principal component analysis (PCA) and fusion were applied on the satellite imagery. Field information collected in September 2018 plus available field data from September 2013 and May 2014, in total 60 sample plots, were used to produce a ground truth map. Forest type was determined through Gorji Bahri approach in each plot. Based on forest types separability, six types were identified (pure beech, mixed beech, beech-hornbeam, mixed hornbeam, pure eastern hornbeam, and eastern hornbeam-Persian oak) to be classified using satellite data. The performance of some classifiers like support vector machine (SVM), random forest (RF), artificial neural network (ANN) and maximum likelihood (ML) was analyzed using two different training datasets.

The results indicated that the sentinel-2 dataset performed better than Landsat-8 for producing forest type map specially when the number of classes increases. It was also found that image fusion methods on sentinel-2 and landsat-8, appropriately improved the result of classifications. This research confirms the effectiveness of number of training samples on the performance of classifiers. Respecting the accuracy assessment criterion, the SVM and RF algorithms showed better result while only 22% of field data was used as training samples. By increasing the number of training samples to 50% of field measurements, the highest accuracy was obtained using RF algorithm applying on all datasets from two satellites.

The Landsat-8 and Sentinel-2 satellite data have moderate capability (overall accuracy around 75% for four-class classification) for mapping forest types in the Hyrcanian forest. The SVM and RF produced more stable and accurate results in comparison with two other algorithms, ANN and ML. Complementary studies are recommended in different forest sites while considering phenology of species and topographic attributes.

In order to determine the spatial pattern of the probability of fire in the forests of Boyerahmad city, Belife evidence function models and support vector machines were used. For this purpose, at first 145 past fire positions were reported, MODIS data and field surveys were recorded using GPS, of which 70% were used for modeling and 30% for model validation. Next, 15 factors (altitude, slope gradient, slope direction, topographic position index, topographic moisture index, surface curvature, distance from village, distance from river, distance from road, geological formations, NDVI, land use, evapotranspiration annual, annual rainfall and annual temperature) were selected to assess the fire risk and maps were prepared. After performing a linear test between the independent variables, the Belife evidence function and the support vector machine models were used to create the fire zoning map. For modeling, past fire locations were identified and 70% of the data collected were used as training data for modeling and 30% for model validation. The results of the fire map study showed that areas with very high sensitivity cover 40% of the area. The results of the validation of the performance of the Belife evidence function models showed that the area under the curve was equal to 72.2% and the support vector machine with the area below the curve was 83%.Predication. Results of the current research can be used to plan and manage future fire hazards in the study area.

Investigation on quantitative characteristics of forest such as Stand volume is one of the most important principles in planning and forest management decision. The aim of this study is comparison of various satellites data capability and non-parametric methods for estimating stand volume of forest. The studied area is district 1 Darabkola forest in Mazandaran province in southeast of Sari with 2612 hectares which is located in 74 basin of Sari natural recourses Department. Using systematic-random with 10 R.sample plots with 300m×500m sampling net system were measured150 circular sample plots. The necessary preprocessing and processing include ratio, vegetation index, Principal Component Analysis and texture analyse were done on WorldView-2¡ Pleiades-2 and IRS-LISS III imagery . For modeling in this study be used different regression methods include different variants of k-Nearest Niebuhr, kernel machine support vector and random forest . The results of modeling the stand volume using machine support vector showed that the best kernel in order for worldview- 2,IRS-LISS III and Pleiades-2 satellites was Polynomial,RBFand Polynomial with %RMSE equal to 34/57,49/5 and 43/03.The best variant in k-Nearest Niebuhr in order for said satellites was chebychev,chebychev and City block with %RMSE equal to 41/18,55/09 and 46/97. %RMSE in random forest method in order for said satellites was 31/33,48/91 and 45/68. Results showed random forest was the best model for estimation stand volume and WorldVeiw-2 satellite data has the best result with percent root mean square error and bias of estimation equal to 31.33 and 2.8 percent.Because of more bands and less width of them, WorldView-2 satellite has better outcomes than Pleiades-2 satellite; since if there are more bands and width of them is narrower, information can be saved in different bands and ratio of signal to noise will be increased. Therefore, phenomenon detects better and accuracy of outcomes increases. The results did not show much difference between the non-parametric algorithms in terms of Percent Root Mean Square Error, but a large difference was observed in terms of sensors. Overall results of this study showed sensors and Regression methods used in this study have a relatively high capability in estimation of forest stand volume . The results also show in addition to the spatial resolution of satellites their spectral resolution has a significant impact on raising the accuracy of the forest stand volume modeling results using satellite images .

Identifying forest areas susceptible to decline in order to take preventive measures can play a significant role in inhibiting this phenomenon. Previous studies suggest the high performance of modeling to identify such areas. Hence, we used 15 models to identify forest areas prone to decline in Lorestan province. For modeling, forest areas with over 50% tree mortality were used as dependent variable, and environmental factors including annual mean rainfall, annual mean temperature, relative humidity, aridity index, evapotranspiration, dust storm index, drought index, distance to surface waters and agricultural lands, slope, aspect and NDVI as independent variables were introduced into the models. AUC of each model was multiplied by its output and the mean of these models was considered as the combined model. The forest decline risk map resulted from the combined model indicated a decline trend from central parts of the Lorestan’s forests to the south and south-western parts. The Random forest and Support vector machine were recognized as the best models with AUC value of 1 and the Bioclim as the weakest model with AUC of 0.75. According to the combined model, approximately 23.7%, 7.5%, and 19.5% of the studied forests had low, medium, and high risk of decline respectively. The climatic factors including aridity index, rainfall, temperature, and evapotranspiration were the most influencing environmental factors, respectively. The present research, in addition to emphasizing the modeling efficiency in identification of forest areas susceptible to decline, indicated that the combination of models yields better result rather than their separate use.

Spatial prediction of fire risk and preparing the forest fire risk map across the natural areas are among the ways that can be used to prevent and to manage fire. The aim of this research was zonation of forest fire risk in Golestan National Park using non-parametric algorithms, namely Artificial Neural Network (ANN), Support Vector Machine (SVM) and Random Forest (RF).
About 100 occurred fire points were considered for modeling the fire risk. The effective factors on fire occurring including vegetation types, physiographic, climatic, and human factors were identified and their relevant maps were prepared from different sources. To modeling purposes, initially the zone was divided into 1-ha levels of decision-making and modeling and then the pixel values of the effective factors on classes of fire occurring, across the 1-ha levels, were extracted and standardized. Based on non-parametric algorithms, fire risk was modeled with 70 percent of the fire points, as training samples. The prepared forest fire risk map was zoned in terms of four classes of low-risk, medium-risk, high-risk and dangerous. The classification accuracy of the maps, resulted from this modeling, was assessed through the overall classification accuracy given 30 percent of the remained fire points.
Results & The results indicated that RF algorithm, with the overall accuracy of 75%, was the best algorithm in predicting the fire risk compare to other ones. Likewise, after matching the fire risk occurring with the results gained from algorithms, it turned out that all algorithms were able to classify the area properly in terms of the fire risk, as more than 80 percent of fire points were placed in the high-risk and dangerous classes