REMOTE SENSING BASED FOREST COVER MAPPING USING RECENT TECHNIQUES

Abstract

Forests are one of the important natural resources found on earth. They are an important component of our ecosystem providing valuable ecosystem services. They produce major forest products such as timber and firewood and various minor forest products such as resin and fibre. They have an important role in conservation of important fauna and flora and myriad to some of the rare plant and animal species. Monitoring and assessment of valuable forest resources is an important component of conservation measures at regional d national level. Increased pressure on forest resources had in recent times had al ad verse impact on other resources such as water, soil and wildlife. At the same time, this as also resulted in climate change causing serious environmental problems. Change in r ee line and shrinkage in snow cover are some of the impacts of forest cover changes in ° cent past. Some of the Himalayan region are not receiving desired amount of snow fall ue to change in forest cover. It is of vital importance for environmental planners and con" rvationists to not only understand the past and present rate of forest change but also be able to predict the future change patterns. For that accurate and precise forest maps can rovide valuable information on the real situation of forest on the ground. The first forest type estimation in India was carried out by two Silviculturists; Champion and Seth way back in 1938. This was further revised in 1968. Today various image processing techniques are used to analyze remote sensing data and generate forest cover maps at various scales. The traditional methods of image classification algorithms are able to produce forest maps to a certain level of accuracy. However, due to their statistical nature and limitations of the algorithm, desired level of accuracy may not be achieved. For accurate mapping of forest cover, high interpretation skills along with advance image processing algorithms is necessary. These algorithms can handle data from various sources producing higher accuracy which the traditional algorithms are unable to handle. The outputs generated using such classifiers are found to produce high accuracy. In recent years, machine learning methods have been used to map forest types as an alternative to statistical classifiers due to their potential for handling non-parametric data easily and generating results of desired accuracy. However, in India, very few attempts have been made to use machine learning methods for forest type mapping. Champion and Seth in their classification used collateral information and intensive field visits for forest type mapping and prediction in those areas where adequate information was not available. With the resources of that time, they Remote sensing Based Forest Cover Mapping Using Recent Techniques brought out the national forest map in 1938 describing 16 major sub-groups spread across the country. The Western Himalayan region which is a myriad of several holy rivers and abundant with rich forest resources, requires precise mapping of forest cover. With the new satellite information and invention of new classification methods, forest cover needs to be mapped with information at individual tree strand level. A review of the literature has been carried out to compare the traditional methods of forest mapping with the advance methods. The studies carried out in India highlight the role of ancillary data in image classification (Singh et al., 2001, Ashutosh et al., 2010). However, the role of ancillary data has been limited either to segment the image based on rule base or to use such data as supportive information in forest mapping. Some studies have tried to use techniques such as neural network for forest density mapping in homogenous forest. Jensen used neural network method to map coniferous forest age using Ian 'sat TM data (1999). In another study use of Support Vector Machines was used to map semi-• 'd vegetation using MISR data on board Terra Satellite. In almost all of the per-pixel studies, 'se of advance classifiers have been carried out for broad based mapping of forest types. Th classification of forest types have been also carried out to map forest groups across b, using AVHRR data (Kennedy and Bertolo, 2002). The use of ancillary data including DEM been successful in mapping of major forest types. It has been observed that in a mountainous region particularly West Himalayan region, the forest types are observed to show a close relationship to their occurrence with the topographical and climatic variables. As such the inclusion of such layers in forest mapping along with the satellite data may be useful for achieving higher accuracy. But the integration of all such layers that are perceived to be affecting forest types may bring out several complexities in creation of datasets and difficulty of handling such huge datasets by the classifiers. Hence selection of variables that have a major bearing on occurrence of forest types becomes necessary. Several statistical methods have been in use to select the useful variables for analysis of field data. However, when the variables are observed to be having multiple relationship, simple statistical techniques may not be able to provide adequate results. Here, techniques such a log linear modeling may be effective for identification of most significant variables and their association that could be effectively used in forest mapping along with remote sensing data. Abstract One of the aims of the present research is to apply different methods of loglinear modeling for identification of suitable variables for further analysis in forest. mapping. After having identified the significant variables, some machine learning and fuzz set based methods for forest type mapping at per-pixel and sub-pixel level have been • attempted. Decision Tree, Evidential Reasoning, Support Vector Machines, Fuzzy c-Means and Artificial Neural Network have been used to generate forest maps at per-pixel level. Whereas Fuzzy Maximum Likelihood Classification, Fuzzy c-Means, Artificial Neural Network and Evidential Reasoning have been used to generate forest maps at sub-pixel level. The results accuracy of classification from these classifiers have been assessed using appropriate accuracy measures. The study area belongs to a part in the Western Himalayan region lying between the geographical extents-31004 07.04" N to 31°06' 07.48"N and 78°04'31.32 E to 78°06'37.46" E. The region has distinct variation of forest types with altitudes. The following datasets have been used to generate various thematic data layers for the study area: a) Remote sensing data: IRS-P6 LISS-III and IRS-P6 LISS IV multispectral and Cartosat-1 PAN images b) Survey of India toposheet at 1:50,000 scale (53JI/4) c) Forest Survey of India Thematic map at 1:50,000 scale prepared from aerial photographs d) Digital elevation model for creation of altitude and aspect layers e) National Bureau of Soil Sample Survey (NBSSS) map for creation of soil color and soil texture map SPSS 16 software has been used for analysis of categorical data. Indigenously developed software have been used for per-pixel and sub-pixel classification of forest types using multisource data. The ERDAS Imagine 9.1 image processing software has been used for processing and for visualizing images. ANN processing has been done in Neural Network Tool Box of MATLAB software for creation of forest type classes using NN. The following thematic data layers have been generated from remote sensing data and other data sources, i) Reference datasets have been generated from IRS-P6 LISS IV, Cartosat-1 PAN images and field inventory data. ii) Multisource data sets have been created using three remote sensing bands and four ancillary layers - altitude, aspect, soil color and soil texture. Remote sensing Based Forest Cover Mapping Using Recent Techniques iii) Agriculture and other non-forest areas have been masked out from the image using high resolution data, topographic sheet and based on field visits. In loglinear model approach, for selection of most significant variables, K-way effects, Partial Association Test and Backward selection methods have been used. The K-way approach has been effective to identify as to which order associations may be helpful in further analysis. A probability value of < 0.05 and high chi-square values was selected as cut-off for selection of good-fit models in all the three methods. The results of partial associations test and backward elimination further helped in identification of significant associations and the major variables useful for forest mapping. Based on the results of the loglinear modeling, the spatial database pertaining to significant variables was reated and used in forest classification process based on various classification algorithms. Five forest types have been recognized and found to be prevailing in the area. Fir and Spruce have been categorized in the same class because of little variation in their leaf structure and tonal variations. The four other classes occurring in the area are-Kail(botanical name-Pinus Wallichiana), Deodar(botanical name-Cedrus Deodara), Chir-Pine(botanical name-Pinus Roxburghii and Banj-oak(botanical name-Quercus Incana). The remaining tree cover that have scanty presence in the area and cannot be mapped separately as a class in the image have been classed as other forest types. For the per-pixel, five non-parametric classifiers have been used. In the fuzzy c-means, three attribute selection norms- Euclidean, Diagonal and Mahalonobis distance have been used. A weighting component of 2 yielded best results for forest classes with accuracy value as high as 70.8% with seven variables. The results of these attribute norms have been analyzed and discussed separately. In Decision Tree Classification, four splitting criteria - Gini Index, Towing Rule, Information Gain and Chi-Square have been used. Overall classification accuracy and user's and producer's accuracy have been separately estimated. An accuracy value of 81.5% has been observed with Towing rule under this classifier. Neural network proved to be a robust classifier and yielded an accuracy value of 80.1% after providing learning rate values and momentum factor obtained after several trials. In evidential reasoning, the value for the parameter "bin size" was varied on subjectively. This in fact is a major limitation of the classifier. [iv] Abstract The use of Evidential Reasoning Classifier and Support Vector Machine methods in forest classification has not been gainful in achieving higher overall accuracy values. A further analysis of individual forest class accuracies was performed so as to evaluate the performance of each classifier in mapping individual forest types. Though ERC could not provide, overall accuracy value more than 60%, yet it was able to map Fir-Spruce with an accuracy 'value of 72.64%. The performance of DT as per-pixel classifier was remarkable in mapping of high altitude conifers-Kail and Deodar and sub-tropical species-Chir-Pine with accuracies values of 92.82%, 90.7 1 % and 84.84% respectively. Among the per-pixel classifiers, Decision Tree Classifier (DTC) and Neural Network (NN) methods gave better results than other classifiers used in the study. The sub-pixel classification approach was used to obtain class fractions within pixels. The reference file generated using high resolution IRS P6-LISS3 and Cartosat-1 PAN image, revealed the existence of class proportions among pixels. The forest type maps generated using different four sub-pixel classifiers reveal that distinct variation of forest type exists with the variation in topographical gradients and soil factors. The sub-pixel classification techniques have also been found to be effective in unmixing the classes within each pixel thereby producing accurate forest area estimates. The highest classification accuracy has been produced from FCM (81.31%) followed by ANN (79.15%), fuzzy MLC (77.18%) and ERC (65.35%). Thus, the research outcome in the form of use of variable selection method based on loglinear model may be quite useful for the forest department at national and regional level, who are continuously collecting voluminous field data and find it difficult to analyze it in the absence of suitable knowledge. The per-pixel and, sub-pixel classification approaches used in the study may be useful in the national forest mapping program of Ministry of Environment and Forests to provide precise and accurate information on forest resources in the country.