Improvement of Forest Canopy Density Mapping of Spare Forests Using a Novel RS-GIS Based Classication Method

Background: Accurate mapping and monitoring canopy cover using remote sensing data as an alternative way for eld surveys are very important for forest managers, particularly in the spare and low dense forests. Due to being area-based of canopy cover density and mixing spectral responses of tree crowns and soil in the thin and semi-dense forests, nding the high-performance method of classication is a challenge particularly on high-resolution imagery. In this study, we compared produced maps of canopy cover using direct remote sensing and indirect (RS-GIS-based) methods in two forest sites on the Quickbird and WorldView-2 images using the Articial Neural Network (ANN) algorithm. Also, the optimal plot area was examined by different plot areas. Results: In the direct method and based on the obtained results, in the Dashte Barm using Quickbird image, the best classication was for plots of 7500 m 2 with an overall accuracy of 56.57% and kappa coecient of 0.32. In the Ilam site and on the WorldView-2 image, the best result is obtained by the plots of 5,000 m 2 area with an overall accuracy of 45.71% and the kappa coecient of 0.263. The results of accuracy assessment of maps of indirect method in the Dashte Barm site for grids with different areas showed that the best classications obtained from sample plot areas of 10000 m 2 with overall accuracy of 82.69% and Kappa coecient of 0.744; but in the Ilam sites the best result was obtained using sample area of 1000 m 2 with overall accuracy of 74.27% and the Kappa coecient of 0.690. Conclusions: The results exposed that use of the RS-GIS based method could considerably improve the results compare to direct classication. Also, the results showed concerning the conditions of canopy cover density of forest stands, plots with different areas can be used to map of forest canopy cover density; however, for direct classication the use of plots with areas of 5000 m 2 and more are suitable in sparse forests. For RS-GIS based method, the plot areas of 1000 m 2 are optimal due to time and cost saving.


Background
Iran is one of the countries with low forest cover, which clearly shows the importance of protecting existing forests. One of the most important vegetation areas in Iran is the Zagros forests, which have long been inhabited by residents and nomads and have been exposed to several damages, which have led to the destruction of parts of these forests. Therefore, the management and monitoring of these forests require a lot of management factors (Afshar 2012).
The forest canopy cover density has a special role in the forest ecosystem and its protection. From a point of ecological view, one of the most important components of the forest ecosystem is the canopy cover (TaheriSarteshnizi 2013), which plays an important role in preserving the forest ecosystem's biodiversity, affecting microclimate and soil characteristics. Forest cover plays a major role in hydrological cycles, temperature systems as well as biochemical cycles. The stability and destruction of vegetation are often a function of the cover and thickness of the canopy so that it can often be Page 3/21 considered as a quantitative and qualitative component of the production of plant communities (Behbahani et al. 2009).
Canopy cover density classes maps are the area-based parameter where the percentage of canopy cover of trees were computed based on the crown area of trees in a certain area such as plots or other suitable de ned polygons by forest managers. However, it should be considered that canopy cover as vertical projections of tree crowns is the sum of the individual crown areas minus their overlapping crown areas (Williams et al., 2003). This area-based parameter can provide useful information for forest management so that preparing the canopy cover density maps is one of the basic information in almost forest management plans particularly in the arid or semi-arid regions with sparse or semi sparse forests. The canopy cover density is generally computed on an area such as topographical-vegetation homogeneous polygons or prede ned knowledge-based grids through the eld surveying or interpretation of aerial photos, which are generally time and cost consuming. One of the alternative sources and ways is using satellite imagery and remote sensing techniques on especially high-resolution imagery ( Mahdavi and Aziz, 2020) using the medium to high-resolution images tried to generate canopy cover density maps in this forest zone, but in the most of studies, researchers could not obtain the favorite results. This unsatis ed result from the use of medium resolution imagery refers to some reasons such as existing the mixed pixels due to be mixing spectral responses of tree cover with ground spectral i.e. bare soil or herbal vegetation at a high dimensions pixel especially in the very sparse forests; as well as to be mixed of spectral responses of two or more different density classes in a pixel especial in the boundary of density classes. It seems that one of the solutions to overcome these mixed pixels is using high-resolution images and classifying images by pixel or object-based classi ers or using novel methods for improvement of the results.
Although the use of the pixel-based classi cation is one of the most used methods for classifying highresolution images (Kim et al. 2011), however, it causes many problems due to being area-based of classi cation objects like forest canopy cover density classes. Due to the presence of shadows and gaps in the sparse forests, and also spectral diversity, structural composition, and heterogeneity in a class, it may lead to errors in differentiating different classes (Bauer and Steinnocher 2001). Since, various methods have been used to solve such problems, such as image pre-processing, such as low-pass lter and texture analysis, contextual classi cation, such as Markov random eld, and post-classi cation processing, such as mode ltering, morphological ltering, and rule-based processing (Kim et al., 2011). Erfanifard et al. (2014) suggested a robust approach to generate canopy cover maps using UltraCam-D derived orthoimagery, which is classi ed by support vector machines algorithm in Zagros forest zone, West Iran.
In addition, the use of object-based methods has recently been used to solve the mentioned problems.
This classi cation uses membership functions to assign each object to the appropriate class (Quynh Trang et al. 2016). In this regard, Rahimizadeh et al. (2020) used the object-based method to study the canopy cover of trees using the Spot-7 satellite image in the high dense forests of Hyrcanian forests (northern Iran) and could improve the results compare to pixel-based classi cation methods. Also, Wang et al. (2004) used the object-based classi cation method on the high-resolution Ikonos imagery to study the mangrove forest cover in the Caribbean coast of Panama on the high dense canopy-covered area.
They stated that the combination of the pixel-based and object-based methods could improve the results.
In general, from these literature reviews on both pixel and object-based methods using medium or highresolution images, it can be found that canopy cover density mapping in the high dense forests has lesser problems compared to the sparse forest like Zagros forests of Iran. However, in the sparse forest regions, we need to nd new suitable methods to overcome these problems.
In line with, Geographical Information Systems (GIS) and its powerful tools assembled on software such as ArcGIS is a computer-based system that capture, prepare, manage, manipulate, analyze, and present geo-referenced data (Ogwankwa 2020). It is assumed that the use of these capabilities can improve the forest canopy cover density mapping in the low dense or sparse forest zones. The GIS and vegetation classi cation systems in conjunction with computer-based automated eld mapping techniques have been prepared the appropriate tools in the process of creating or up-to-dating the vegetation maps (Ismail 2010). In a research which was done on the Zagros forest (Erfanifard et al., 2014), the capability of GIS tools is used to map the forest canopy cover density classes as a robust or novelty method to improve the classi cation results.
On the other hand, choosing an appropriate plot area for forest canopy cover density mapping is very critical due to the spatial variation of forest canopy cover classes in the Zagros forest zone. So, some researchers have done some studies (Adeli et al., 2008;Mahdavi and Aziz, 2020) in semi-arid forest zones to nd the best plot area for eld canopy cover surveys. For instance, Adeli et al. (2008) stated and suggested the different sample areas for different forest canopy cover density classes. Since the plot area can affect forest canopy cover estimations both in eld surveys and remote sensing surveys; nding an optimal plot area is very important particularly when using high-resolution images.
Given the above and the importance of the canopy in forest resource management, it is important to provide a method that can provide quality information in a short time. Therefore, in this study, Quickbird and WorldView-2 images were classi ed using the direct (plot-based classi cation) and RS/GIS-based (indirect) method, which is an experimental and new method, and the canopy cover map of trees in two the area was extracted from the forests of Zagros. The main purpose of this study is to provide a new way to provide higher-quality forest canopy cover density maps. Besides, nding the optimal plot area for forest canopy cover mapping in sparse forests like Zagros forest using high-resolution images is the second propose in this research.

Study areas
The study areas were located in two sites in two different spare and semi-dense forest zone. These sites were selected based on the availability of two high-resolution satellite data sets i.e. Quickbird (Dashte Barm) and WorldView-2 (Ilam) that were purchased and used in other studies (Fig. 1).

Ground truth data
The forest canopy cover density maps were prepared on sample plots where distributed and selected randomly using a combination of eld survey, visual and digital interpretation of fused satellite images trough fusion of bands (multispectral and panchromatic bands) of Quickbird and WorldView-2 images, as well as the use of Google earth images in both sites. Totally, 52 quadrangular concentric plots for the Dashte Barm and 60 quadrangular concentric plots for the Ilam were designed and implemented randomly in the eld (Fig. 3) with areas of 1000, 1500, 2500, 5000, 7500 and 10000 m 2 . The percentages of canopy cover of plots were computed based on measuring the crown area of whole trees in each quadrangle. The percent of canopy cover area was classi ed into ve classes including very thin (1-10%), thin (10-25%), semi-dense (25-50%), massive (50-75%) and very massive (75-100%) based on the Iranian FRWMO guideline. It should be considered that the class of very massive does not exist on both sites. Then, 70% of the plots from each class were considered for training samples, and 30% of the rest samples were selected for validation of the results. Figure 2 shows the spatial distribution of plot centers.
Remote sensing data In this study, the high-resolution images of Quickbird (Dasht Barm site) and WorldView-2 (Ilam site) were used for classi cation (Fig. 3). The Quickbird satellite images are composing four color bands (blue, green, red, and near-infrared), and a panchromatic band; and WorldView-2 images comprising eight color bands and a panchromatic band (see Table 1). Image processing Data preprocessing is one of the most important steps in image processing. The type and manner of performing this operation depend on factors such as the type of data used and the purpose of the research (Saraskanrood et al. 2019). Both used images were orthorecti ed using enough ground control points and DEM of the study area. Then, to accurate extract tree crowns, multispectral and panchromatic bands were pan sharpened using the Brovey method, so that the ne resolution multispectral images with ne extraction of tree crowns. In addition, some important vegetation indices of NDVI, GNDVI, and SAVI were also created ( Table 2) and used together with other bands in the classi cation.

Classi cation
The rst step for classi cation images is the de nition and selection of class thresholds according to the characteristics of the study area and satellite images (Boyaci et al. 2017) as well as based on guidelines of responsible organizations. Therefore, ve classes based on the guidelines of FRWO were considered to prepare a canopy cover map in the study areas. For the investigation on the optimal plot area, classi cations were done with Quadrangular areas of 1000, 1500, 2500, 5000, 75000 and 10000 m 2 . The classi cations were performed using an Arti cial Neural Network (ANN) algorithm. This algorithm is an information processing system composed of the structure and function of the physiological human brain neural network and some theoretical abstraction, simpli cation, and simulation of several basic characteristics (Xu et al. 2020). The images were classi ed by two methods, including direct and indirect (RS-GIS based) methods.

Direct method
In this method, the classi cation of images (pan-sharpened bands together with vegetation indices) was done using Arti cial Neural Network (ANN) algorithm and the function Sigmoid or Logistic Activation Function (Fig. 4) by 70 percent of samples taken in each canopy cover class by eld surveys and Google earth images. All sigmoid functions have the property that they map the entire number line into a small range such as between 0 and 1 (Bonnell 2011). The classi cations were repeated for different sample areas. The accuracy assessment of the classi ed maps was evaluated using test samples (30%).

Indirect (RS-GIS based) method
In this method, rstly, the tree and non-tree objects (i.e. bare soil and roads) were separated through the classi cation of the pan-sharpened images (Brovy method) together with vegetation indices using the ANN algorithm and Sigmoid or Logistic Activation Function. Table 3 shows the number of training pixels for each class. The accuracy of classi cation results was assessed by test samples. Then, using ArcGIS software tools, the images were reclassi ed into two classes of trees, and non-trees by merging bare land and roads; then the reclassi ed image rasters were converted to polygon vector shape les. The study areas were gridded into 1,000; 1,500; 2,500; 5,000; 7,500 and 10,000 m 2 polygonal grids, and the percentage of canopy cover was calculated in each polygonal grid based on the sum of tree crown area in each grid by intersection function. In the next step, the canopy cover density maps were prepared based on canopy cover density classes. The polygonal canopy cover density classes were again converted to raster format for accuracy assessment. The accuracy assessment of obtained maps was done using test samples. It should be noted that the test samples which are used to evaluate the classi cation results, were the same ones that were used in the direct classi cation method to avoid uncertainties.

Direct-method
In the direct method, the forest canopy cover density map was obtained using the classi cation of satellite images with areas of 1000, 1500, 2500, 5000, 7500, and 10000 m 2 and ANN algorithm.
According to the obtained results (Table 4), in the Dashte Barm site using Quickbird image, the best classi cation is related to 7500 m 2 plots with an overall accuracy of 56.57% and kappa coe cient of 0.32 ( Fig. 5-a) and the lowest result came from 1000 m 2 samples (overall accuracy of 48.15% with a kappa coe cient of 0.247). Since the Dashte Barm site is a very spare and low-cover forest area, only three forest canopy cover density classes were classi ed using the ANN algorithm. Also, in the Ilam site and on the WorldView-2 image, the best result is obtained by the samples of 5,000 m 2 area with an overall accuracy of 45.71% and the kappa coe cient of 0.263 (Fig. 5-b). The lowest result belongs to the 1000 m 2 samples (overall accuracy of 38.13% with a kappa coe cient of 0.160). In this site (Ilam), the amount of canopy is richer and there are classes with a high percentage of canopy cover, but the very dense class has not been seen in the study area.

Indirect method (RS-GIS based method)
In this method, rst, the pan-sharpened images were classi ed into different classes (trees, roads, and bare land) using point training samples and then roads and bare land classes were merged. The results of the accuracy assessment of the classi ed images can be seen in Table 5. In the second step, each image was divided into 1000, 1500, 2500, 5000, 75000, and 10000 m 2 grids, and the percentage of tree canopy cover for each grid was calculated based on areas of tree polygons and area of grids. Figure 6 shows the tree crown area polygons obtained from two-class classi cation of Quickbird images and the grids with different areas.
The results of accuracy assessment of canopy cover maps in the Dashte Barm site for grids with different areas showed that the best classi cations obtained from sample plot areas of 10000 m 2 (overall accuracy of 82.69% and Kappa coe cient of 0.744) (Fig. 7-a); but in the Ilam sites the best result was using sample area of 1000 m 2 (overall accuracy of 74.27% and the Kappa coe cient of 0.690) ( Fig. 7-b) (Table 6). Optimal plot area The optimal plot area for measuring and classifying the canopy cover of trees was investigated for two methods in this study. In the direct classi cation method, the area of the plots did not have a signi cant effect on the classi cation results. However, in the Dashte Barm site (Quickbird image), using the plot area of 5,000 m 2 , and the Ilam site (WorldView-2 image), the plot area of 10,000 m 2 could slightly improve the classi cation results. In the RS-GIS-based (indirect) method, in the Dashte Barm site (Quickbird image), using the 10,000 m 2 plot area, and in the Ilam site (WorldView-2 image), using the 1,000 m 2 plots provided better results for classi cation (Fig. 8).

Discussion
In this study, the forest canopy cover density of two forest sites was mapped using direct image classi cation and indirect classi cation (integration of RS and GIS) methods. It is the traditional way to use a direct-method in classifying forest areas using satellite images through plot-based classi cation with a moderate accuracy result. However, to improve the classi cation results, we tried to use an indirect method, which is an integration of tree-based remote sensing (RS) classi cation, and then using the geographical information system (GIS) capability tools for extracting the tree crowns and then computing the plot-based canopy cover density classes. We also tried to nd the most appropriate plot area for forest canopy cover mapping in both direct and indirect methods with different areas.
According to the obtained results, it was found that in the direct method, the forest canopy cover map does not have high accuracy and the best result was 57.39% for the Dashte Barm site (Quickbird) and 46.27% for the Ilam site (WorldView-2). As can be seen in Table 4, the classi cation results for both sites are not of acceptable accuracy. The main reason is the mixing of the spectral responses of vegetation and soil as well as the variety of the tree covers densities in Zagros regions. Since, our the study areas are sparse and semi-dense forest and; so in a plot-based classi cation, due to the spectral interference of vegetations dominantly trees with soil, separating different canopy cover density classes from each other will be with wrongs so that it is not possible to classify images with desired classi cations with high accuracy . Although, in this study, the SAVI vegetation index was used to reduce the effect of soil and increases the re ectivity of plants with a low percentage of cover (Imani et  In the arid and semi-arid forest zones where have sparse and semi-dense forest cover, the literature reviews of last previous studies showed that in the plot-based classi cation of forest canopy cover density mapping, using high-resolution images compare to medium resolutions cannot be more effective to considerable improvement. Of course, it is necessary to pay attention to the fact that in areas similar to the studied forests, more images with low resolution and relatively high have been used to prepare the canopy cover of trees, the results of which also show there has been insu cient accuracy of these images (Boyaci et al., 2017). However, in a study conducted in forests with high vegetation cover, Banerjee et al. (2014) and Rahimizadeh et al. (2020) presented the canopy cover classi cation of trees with high accuracy using a direct-method using the Landsat-TM and Spot-7 satellite images, respectively.
It should be also considered that in Iran with about 14 million hectares of forest, the dense forests covering only 12.4% of the total forest area, and the rest of the forest areas have semi-dense and open canopy cover (FRWMO, 2020). Therefore, paying attention to these forests and obtaining accurate information about the canopy cover rate and its percent is of great importance for their current and future management. Therefore, using techniques and methods such as the combination of remote sensing (RS) and geographical Information Systems (GIS) are necessary to improve the results. The idea of separating the accurate boundary of tree crowns on very high-resolution satellite images as pan-sharpened with a resolution of lower than 0.5 meters or aerial digital photos taken by airplanes or UAV drones and then considering the crown areas in a plot to computing canopy cover densities is a way to overcome the problems of previous methods.
The results of the RS-GIS based method for classifying the forest canopy cover density have shown a considerable improvement compared to the direct-method in both sites. As can be seen in Table 6, the results of this method have higher overall accuracies than the highest classi cation results using the direct-method. Overall, the RS-GIS-based method could improve the classi cation results by an average of 23.25% for the Quickbird image and 17.69% for the WorldView-2 image so that overall accuracy of about 83 percent and Kappa coe cient of 0.74 were obtained in the Ilam site. According to the results of Chuang et al. (2011), if the overall accuracy is more than 70%, the accuracy of production drawings is reliable. Therefore, the use of the RS-GIS method provided satisfactory results for canopy cover mapping.
The use of remote sensing technology is important due to lower nancial costs and more time savings (Saraskanrood et al., 2019). Therefore, the issue that was considered in this study was the appropriate time to prepare the intended maps. Of course, the more accurate the information used in a shorter period, the more valuable that method is. In the methods studied in this study, it was found that the direct classi cation method, according to the area of training samples and the number of effective pixels in the classi cation, processing time can be variable and the larger the area of training samples, so the canopy classi cation time of trees increases. Therefore, in using images with high spatial resolution on a larger scale, powerful and high-performance systems are needed. However, this problem was solved in the RS-GIS based classi cation method, and after the initial classi cation, which was done to separate the crown area of the trees, the rest of the steps were done easily, faster, and with less time (less than 1 hour). It should be noted that the mastery and experience of the researcher can be effective during processing.
Also in this study, the most appropriate plot area was presented to achieve the highest classi cation accuracy. In forest inventories, selecting the optimal sample area must be done for su ciently and accurate gathering information with considering accuracy and time-cost factors (Mirzaei et al. 2014). As the results showed, in both used methods, with changing the plot area, the classi cation results also will change depending on forest canopy condition.
As can be seen in Fig. 8, in the direct-method, the graphs show a speci c trend from 1000 to 10000 m 2 plot areas so that with increasing the plot area, the accuracies increased so that classi cations with the 5000 m2 and 10,000 m2 plot areas have high performance. In conclusion for direct classi cations, the area of larger plots has been able to improve the classi cation results, however, according to the FAO (2000) de nition, which is considered the areas of more than 0.5 as one of the factors for the de nition of forest area, so use it makes more sense than plots larger than 5,000 m 2 .
In the indirect (RS-GIS based) method, rst, the trend of the diagram is downward and then is upward so that with increasing of plot area from 1000 m 2 to 2500 m 2 , the accuracies will decrease but from 2500 m 2 to 10000 m 2 , the result accuracies are increased. However, in the Ilam site and WorldView-2 images, plots with 1000 m 2 have the best results. These differences of accuracies in the two sites refer to different variety of forest canopy cover density in the sites. It seems that in forests with a variety of forest canopy cover classes, due to be more number of classes as well as to be near to spectral responses of canopy cover density classes, separation of classes is very di cult.

Conclusion
In general, the results show that the combination of remote sensing techniques and GIS tools can improve the results of forest canopy cover classi cation in both study areas. Therefore, separating the boundaries of tree crowns and computing these areas for canopy cover densities using GIS tools is a new method in future research in the arid and semi-arid zones with sparse forests. In this regard, the use of high-resolution images can have a positive effect on improving the results. Also, the results showed that concerning the conditions of canopy cover density of forest stands and in the forest regions with thin and semi-thin forest canopy cover densities such as Dahte Barm site, the use of plots with areas of 10000 m 2 and more are suitable for direct classi cation. But, for RS-GIS based classi cations, the plot areas of 1000 m 2 are optimal due to time and cost saving. However, in the forest regions with semi-dense to severity dense canopy cover classes such as the Ilam site, using 1000 m2 plot areas is enough and optimal. The results also showed that the forest canopy cover density map of Dashte Barm site using Forest canopy cover density maps by 7500 m2 plot areas at the Dashte Barm site (a), and 5000 m2 plot areas at the Ilam site (b).

Figure 7
Tree canopy cover map, (a) using a 10,000 m2 grid in the Dashte Barm site (Quickbird image), (b) Tree canopy cover map using a 1000 m2 grid in the Ilam site (WorldView-2 image).