Essay Writing Service

A Spatialized Classification Approach For Land Cover Mapping Using Hyperspatial Imagery

do not necessarily reflect the views of


Maps of classified surface features are a key output from remote sensing. Conventional methods of pixel-based classification label each pixel independently by considering only a pixel’s spectral properties. While these purely spectral-based techniques may be applicable to many medium and coarse-scale remote sensing analyses, they may become less appropriate when applied to high spatial resolution imagery in which the pixels are smaller than the objects to be classified. At this scale, there is often higher intra-class spectral heterogeneity than inter-class spectral heterogeneity, leading to difficulties in using purely spectral-based classifications. A solution to these issues is to use not only a pixel’s spectral characteristics but also its spatial characteristics. In this study, we develop a generalizable spatialized classification approach for high spatial resolution image classification. We apply the proposed approach to map vegetation growth forms such as trees, shrubs, and herbs in a forested ecosystem in the Sierra Nevada Mountains.  Our results founds that the spatialized classification approach outperformed spectral-only approaches for all cover classes examined, with the largest improvements being in discriminating vegetation classes.

1. Introduction

One of the most important applications of remote sensing is to use imagery to classify and delineate different objects and land cover types on the earth’s surface. Historically, remote sensing has been used to perform land use/land cover (“LU/LC”) mapping, in which complex mixtures of vegetation and non-vegetated surfaces are collapsed into a relatively small number of discrete classes labeled according to, typically, their relative fractional cover. For instance, the National Land Cover Database (“NLCD”) 2011 Classification System (Jin et al., 2013) defines a “forest” as an area of land with greater than 20% tree cover. While medium and coarse scale remote sensing (e.g. Landsat and MODIS imagery) with their relatively high temporal frequency are extremely valuable for various long-term research and management objectives such as long term sustainable forest management, biodiversity monitoring, carbon accounting, habitat protection, and sustainable timber production, these sensors lack the spatial detail to resolve fine landscape features such as individuals trees and shrubs, and are thus unable to produce many of the inventory products that are necessary for a full understanding of ecosystems processes (Falkowski, et al., 2009; Pu & Landry, 2012). The impact of the focus on LU/LC mapping can be significant; the integration of these discrete LU/LC classes into ecosystem modeling versus using continuous fractional covers of plant functional types leads to profound differences in modeled energy and water flows (Bonan, Levis, Kergoat, & Oleson, 2002). Clearly, there is a need for remote sensing analyses to go beyond producing discrete LU/LC classes, and move towards an inventory-based approach to monitoring ecosystem characteristics, analogous to what field inventories can produce.

“Hyperspatial” remote sensing, defined as image data with pixels smaller than some of the objects of interest (for trees, typically ≤ 1 meter ground sample distance, “GSD”), has been employed in monitoring and obtaining forest inventories data at an individual plant scale, including fractional plant functional type cover, crown size, species, and aboveground biomass (Gougeon & Leckie, 2006; Greenberg, et al., 2006; Greenberg, et al., 2005; Key, 2001).  Historically, aerial imagery comprised the bulk of this type of image data, and the analysis techniques relied on significant field assessments and manual interpretation of the imagery. With the increasingly availability of hyperspatial satellite imagery as well as the acceleration of unmanned aerial system (“UAS”)-based inventorying, and the advance of computing infrastructure and computing power, there is a general trend of obtaining and updating long-term detailed forest data via semi-automatic or automatic processing of high spatial resolution imagery (Gougeon & Leckie, 2006).

Although hyperspatial imagery often outcompetes medium and coarse resolution satellite imagery for the level of details that can be observed and extracted from the earth’s surface, new challenges have emerged for detailed vegetation classification using high spatial resolution imagery. Some of the challenges reported in the literature include how to deal with shadows caused by trees, high intra-class spectral variations, and high inter-class spectral homogeneity (Lu & Weng, 2007). Conventional classification techniques used with medium and coarse spatial resolution imagery focus on purely spectral-based classifications, but these techniques do not appear to work in many circumstances when applied to hyperspatial imagery (Myint et al, 2011; Puissant et al., 2005). One likely explanation for this is that at medium to coarse scale resolution (typically >1m GSD), each discrete LU/LC class tends to have a separable spectral reflectance signature through the electromagnetic spectrum that is captured by different bands in the imagery. At this scale, the intra-class spectral variations often cluster around the ideal spectral reflectance signature distribution and can be described by a probability distribution to model the classification based purely on the spectral characteristics. At high spatial resolution, however, the pixel sizes can be smaller than classes of interest, and there exists high spectral variations within a given class as well as among classes. The weak separability of spectral characteristics often makes it impossible to use solely spectral information to achieve satisfactory classification results. Using conventional spectral-based classification techniques on hyperspatial imagery results in pixels with identical spectral responses belonging to different classes being misclassified, and classification results often have a “salt-and-pepper” effect, where pixels falling within a single object (e.g. a tree) may be classified into multiple different classes. The limitations and decreasing of classification accuracy when applying conventional pixel-based approaches in high spatial resolution image classification have been stated in several studies (Greenberg et al., 2006; Lu & Weng, 2007; Marceau et al., 1990; Woodcock & Strahler, 1987; Yu et al., 2006).

Spatial information has been incorporated into hyperspatial image classification in a number of different ways including 1) the inclusion of texture features in a classifier, calculated as summary statistics (e.g. mean, variance, kurtosis) from fixed-sized local windows around a given pixel, 2) object oriented analysis (“OOA”) in which an image is pre-segmented into polygons (“objects”) of homogenous spectral characteristics, and these polygons become the target of classification, often using spectral and textural information falling within the polygons as predictors (Hamada et al., 2011; Kim et al., 2009; Lu & Weng, 2007; Myint et al., 2011; Yu et al., 2006), 3) custom object detection using computer vision techniques, in which algorithms that are geared toward a specific type of land cover class are applied, e.g. individual tree crown detection (Falkowski et al., 2009) or road detection (Valero et al., 2010).  These algorithms largely suffer from similar problems, in that they typically rely on significant assumptions about the spatial/spectral characteristics of the land cover classes that may, in fact, not be present in reality; in other words: they often lack generalizability.  Texture windows, for instance, require the apriori determination of the size and shape of the window(s) to be used in the classifier, and these choices can significantly impact the classification, as well as effectively reducing the resolution of the final product to the scale of the window used.  OOA approaches assume a unique object is, more or less, spectrally homogenous which, in the case of, for example, a tree crown with a combination of sunlit and shaded pixels, is not a valid assumption.  Custom object detections can work well with specific types of land cover classes, but lack generalizability; i.e. each different type of object may need its own custom detection approach.

We propose a new approach that uses spatial and spectral information from a region surrounding a pixel to be used in a classifier.  Our approach does so in a generalizable way: no apriori assumptions are required as to the spatial patterns of a given class, the consistency of these patterns from one class to another, or the homogeneity of the spectral values.  Herein, we develop the theory of the spatialized classification method and provide a case study towards a land cover classification of hyperspatial imagery for a mixed conifer forest in the Sierra Nevada Mountains (California/Nevada).

2. Methods

2.1. Theory

Let X be a three-dimensional image array with row and column dimensions I and J, and N layers.  Xi,j,n is a single cell location in X  at row/column i,j and layer n.  The goal of a statistical classifier is to produce an image Y with the same dimensions as X in which each cell Yi,j is a unique class.  The conventional “spectral-only” approach to classification determines the class of a pixel based only on that pixel’s spectral values:

Yi,j = f({Xi,j,n, 1 ≤ n ≤ N})

Spatialized classification expands the set of predictor variables to include not only the layer values, but neighboring pixels as well.  Given a maximum search window radius of R cells, the training dataset for the spatialized classification approach becomes:

Yi,j = f({Xi,j,n, i-R ≤i≤i+R, j-R≤j≤j+R, 1 ≤ n ≤ N})

A single predictor variable, then, represents the pixel value at a fixed, relative distance from the pixel to be classified.  This approach dramatically increases the number of predictor variables P available for use in a classifier from P = N to P = (2R+1)2 x N.  For example, with an 8-band image (N=8) and a window radius of 3 (R=3), the number of predictors for use in a spectral-only classification would be P = 8, whereas in the spatialized classification would be P = 392.  We note that unlike a texture feature analysis, this approach does not summarize pixel values; instead, it uses all of the raw pixel values in the local window.

A spatialized training dataset almost certainly contains a large amount of useless or redundant predictors. Proximate pixels within the local window may share similar characteristics and, due to a high correlation, may not add to the predictive power of the classifier.  More distant pixels from the center pixel may not be relevant to the identification of smaller objects (e.g. if a tree has a crown diameter of 2 meters, a pixel 20 meters away may not provide useful information).  In addition to redundant/useless predictors, as the search radius R increases, the computational costs to both train a classifier, as well as apply the classifier to an image increase exponentially.  Thus, the specific algorithm used with the spatialized classification training data should have the following characteristics: 1) must be able to use overpredictive datasets without overfitting the classes, 2) should be able to identify the predictor importance so the search radius can be potentially reduced, improving computational efficiency.  These characteristics are common with machine learning algorithms such as Random Forests (Breiman, 2001), Support Vector Machines (Melgani & Bruzzone, 2004) and Neural Networks (Benediktsson, Swain, & Ersoy, 1990).

2.2. Implementation

To demonstrate the spatialized classification approach, we performed a classification of vegetation lifeforms found within forests across the eastern side of the Lake Tahoe Basin (California/Nevada) using hyperspatial Worldview-2 imagery using both our spatialized classification approach, and a spectral-only approach for comparison.  Our target classes are described in Table 1.

Table 1: Target classes and number of training and testing pixels.

Class Dominant Species

Number of training/testing pixels

Conifer Tree Abies concolor, Abies magnifica, Calocedrus decurrens, Pinus contorta, Pinus jeffreyi


Hardwood Tree Populus tremuloides


Tall Shrubs Alnus tenuifolia, Salix spp.


Shrubs Arctostaphylos nevadensis, Ceanothus cordulatus, Ceanothus prostrates, Chrysolepis sempervirens, Quercus vaciniifolia, Spiraea densiflora ssp. splendens              23,719





Barren and Impermeable




2.2.1. Study Area, Image Data, Field Data, and Ancillary Datasets

The study area is the northeastern side of Lake Tahoe Basin, located in the Sierra Nevada Mountains along the border of California and Nevada. Elevation ranges from 1,900m to 3,050m above sea level (a.s.l.). The climate of the Lake Tahoe Basin follows a Mediterranean pattern with long, cool wet winters and short, warm dry summers. Precipitation usually occurs between October and May as snow. The topographic complexity leads to high variations in temperature, precipitation, and solar radiation and has resulted in a rich diversity of vegetation types. The area is dominated by a variety of conifer species including White Fir (Abies concolor), Jeffrey Pine (Pinus jeffreyi), Red Fir (Abies magnifica), Lodgepole Pine (Pinus contorta), and Incense Cedar (Calocedrus decurrens). Broadleaf tree and tall shrub species are also present, and include Quaking Aspen (Populus tremuloides), Mountain Alder (Alnus tenuifolia) and several species of Willows (Salix spp.).  In addition to trees and tall shrubs, many species of shorter shrubs and herbs are present (Barbour et al., 2007).

We used WorldView-2 (“WV2”) imagery in this analysis. Worldview-2 is an 8-band, multispectral imager that collects data at 0.45 m across a wide panchromatic band, and at 1.85m for the 8 visible/near-infrared spectral bands that range between 400 to 1040nm. The acquisition was acquired on September 10th, 2010. The imagery was collected at an off-nadir angle of 8.39°, with a solar azimuth of 137° (northwest) and a solar elevation of 54.22°. The imagery covers a total area of 422.48 km2. The imagery was atmospherically corrected and orthorectified by the USDA Forest Service Region 5 Remote Sensing Laboratory. The multispectral imagery was pansharpened to 0.45 m using PCI Geomatica’s PANSHARP algorithm (Zhang, 2002).

To use in supporting the production of training as well as the independent validation datasets, we utilized field data collected by the USDA Forest Service and the University of California, Davis.  This data was created by choosing random plot centroids falling within the Lake Tahoe Basin, CA/NV, and then creating a regular grid centered around those random points.  These points were loaded into an iPad connected to a Bad Elf GPS and running the CartoMobile field GIS software with the Worldview-2 image as the background layer.  Using both the GPS location as well as in-field photointerpretation, each of the points were identified down to species (if a tree), or to general land cover material if some other type of cover.  In addition to the raw imagery and field data, we also utilized Google Earth historical imagery, Google Street View, and the Tahoe Basin Exiting Vegetation Map (TBEVM) (Greenberg et al., 2006) to increase our training/testing sample size.  In total, our training and testing dataset included 247,648 pixels.  We performed stratified random sampling on this dataset based on the target classes using a 50% split between training and testing datasets.

2.2.2. Classification Algorithm: Random Forests

We chose to use the machine learning algorithm Random Forests (Breiman, 2001) as our classifier.  Random Forests is an ensemble decision tree based classifier, where each tree is trained using a bootstrap sample of m from the original training samples M and each split occurs at a variable chosen from a random subset of h from the original variables Η (Breiman, 2001). For classification, the final result is based on a majority vote over all of the trees (the “Forest”). By using a newly sampled random subset of variables for splits in each tree, the correlations between each tree in the ensemble are reduced and computational time is saved. Since the subset of variables chosen to be used in each tree is different and variables being sampled to a set can be in a new set next time, this helps stabilize the classification accuracy and acquire generalizable classification results where a small change in the nature of the training samples will not dramatically alter the classification accuracy (Breiman, 2001).

For this analysis, we utilized the RandomForestSRC implementation of Random Forest (Ishwaran et al., 2008) within the R Statistical Computing framework (R version 3.1.3, 2015).  This implementation had several improvements over the original implementation, namely 1) native support for parallel processing during both the model construction as well as the model prediction phase, 2) a memory-safe and more efficient variable importance algorithm (VIMP) known as the “Maximal Subtree” method (Ishwaran et al., 2010).

An issue that has been identified with highly overpredictive datasets when used with a Random Forests classifier is the impact of inter-class sample size imbalances in the training dataset (Khoshgoftaar et al., 2007).  Without a balanced number of samples per class, Random Forests often need significantly more trees to produce a stable classification.  As such, we implemented a downsampling approach to balancing the training dataset before it was passed to the Random Forests algorithm: given the minority class’s sample size D, we randomly removed training samples from the majority classes to reduce each of their sample sizes to D.

To speed up the prediction, the Random Forests models that were produced went through two optimization phases.  First, the model was generated with all of the downsampled training samples, and all of the variables.  The first optimization that was employed used the Maximal Subtree method (Ishwaran et al., 2010) to determine the subset of predictor variables that were considered important to the model.  The model was then re-run using only these variables.  Next, the change in error as each additional tree was added was examined, and the number of trees at which the error rate improved by less than 0.1% was determined.  The model was then re-run using this reduced number of trees.  This final model was the optimized model used in the prediction on the image datasets.

2.2.3. Iterative Training Process

The training process followed an iterative procedure as follows:

A set of random image subsets were extracted falling within each of the unique FGDC subclasses as identified by the TBEVM map (Greenberg et al., 2006).

Pixels belonging to the dominant lifeform/cover classes for each of the image subsets were photointerpreted and digitized, guided by the USDA Forest Service/UC Davis field data, Google Earth historical imagery, and Google Street View.  For example, for an image subset falling within an FGDC subclass identified as an “Evergreen closed tree canopy”, pixels identified as evergreen (conifer) trees would be digitized.

The pixels from the local window surrounding each of the identified training pixels were extracted out to a radius of 7 pixels (a 15 x 15 window, 6.75m x 6.75m).

The training data was used to generate a Random Forest model.  Downsampling to the minority class size was used to balance the training data, and the model was optimized using the Maximal Subtree method and the minimum-needed-trees.

The model was applied to a new set of randomly chosen image subsets following the process in step #1.  This produces a classification product.

The image subsets were examined for errors.  Pixels that were incorrectly classified in step #5 were identified and added to the existing training data locations with the correct class label.

Steps 3-6 were repeated until the predictions appeared to stabilize.

We repeated this process multiple times, yielding a total of 237,649 training samples.  Table 1 shows the final number of training samples for each of the target classes that were identified through photointerpretation.

2.2.4. Determination of Maximum Search Radius

The only real parameter that is required by this method is the maximum search radius.  This value must balance accuracy with computational efficiency.  As such, we ran the model on several versions of the training data, in which only features from within a given distance from the center pixel were used, ranging from 1 to 7 pixels in radius.  At each window size, the number of total available variables and selected variables were recorded, and percentage of important variables chosen from all available variables was calculated.  Figure 1 shows that until the window size exceeds 7 x 7 pixels, the model uses all available pixels within a window.  A window diameter of 15 x 15 pixels (6.75m x 6.75m) was found to be sufficiently large to capture all of the important variables.

Figure 1: Percent of variables picked for various window sizes.  This plot is used to set the optimal window size.

2.2.5. Accuracy Assessment and Variable Importance

Once the final model was constructed, it was applied to an independent set of validation data that was not used to construct the models.  Table 1 shows the per-class sample sizes of the validation dataset.  We produced a confusion matrix, and calculated the overall accuracy, kappa coefficient, and User’s and Producer’s accuracies.  We also applied the model to a set of image subsets with differing vegetation structure and composition for a qualitative (visual) analysis of the results.  In addition to the accuracies, we performed an analysis of the variable importance (the subset of variables chosen for the final model) and produced histograms of the distance from center, azimuth from center, and bands chosen.

2.2.6. Comparison with single-pixel approach.

To compare against the classic, non-spatial approach, we re-ran the model as before with the same training dataset, variable reduction, and tree optimization, but only using the center pixels as predictors, e.g., we did not use any spatial data in the classifier.  We calculated a confusion matrix, the overall accuracy, kappa coefficient, and User’s and Producer’s accuracies for the single-pixel classifier.

3. Results

The final model results in an overall classification accuracy of 96.0% with a Kappa of 0.95.  User’s accuracy ranged between 74.1% (conifer trees) to 100% (water), with hardwood trees being the most accurate of the vegetation classes (96.1%).  By comparison, the non-spatialized approach (spectral-only) resulted in an overall classification accuracy of 89.9% with a Kappa of 0.87.  User’s accuracy ranged between 32.0% (short shrubs) to 100% (water), with hardwood trees being the most accurate of the vegetation classes (89.8%).  Across all classes, the spatialized classification outperformed the spectral-only approach.  Table 2 shows the confusion matrices and Table 3 the per-class User’s and Producer’s accuracies for both approaches.

Table 2: Confusion matrix of target classes for a) spatialized classification approach and b) spectral-only approach.

a. Spatialized classification Barren and


Conifer Tree

Herb Hardwood Tree

Soil Shrub Tall Shrub

Water # of

Classified Pixels

Barren and Impermeable 11177









Conifer Tree 40









Herb 0









Hardwood Tree 0









Soil 51









Shrub 13









Tall Shrub 0









Water 0









# of Ground Truth Pixels 11281









b. Spectral-only classification Barren and


Conifer Tree Herb Hardwood Tree

Soil Shrub Tall Shrub

Water # of

Classified Pixels

Barren and Impermeable 11025 0 30 1 58 0 0 268 11382

Conifer Tree 1 3270 116 3565 0 66 262 0 7280

Herb 0 89 19513 192 2 90 5 0 19891

Hardwood Tree 0 946 143 15970 0 86 648 0 17793

Soil 226 4 29 6 5714 8 0 11 5998

Shrub 18 407 902 1629 50 1458 82 8 4554

Tall Shrub 0 200 3 2356 0 31 5572 0 8162

Water 11 0 0 0 1 0 0 48752 48764

# of Ground Truth Pixels 11281 4916 20736 23719 5825 1739 6569 49039 123824

Table 3: User’s and Producer’s accuracy of target classes using spatialized and spectral-only classification.

Spatialized Classification Spectral-Only Classification

Class Producer’s

Accuracy User’s

Accuracy Producer’s

Accuracy User’s


Barren and Impermeable 99.1%


97.7% 96.9%

Conifer Tree 89.8%


66.5% 44.9%

Herb 97.7%


94.1% 98.1%

Hardwood Tree 85.5%


67.3% 89.8%

Soil 99.6%


98.1% 95.3%

Shrub 98.0%


83.8% 32.0%

Tall Shrub 95.0%


84.8% 68.3%

Water 99.9%

100.0% 99.4% 100.0%

Figures 2 through 4 show good visual agreement with the predicted classes.  Figure 2 shows regions of relatively open conifer forest.  The conifer trees, shrubs and soil in the open canopy areas were captured by the model and classified properly.  Figure 2 shows image subsets containing mixed open conifer forest with shrub ground cover, and Figure 3 shows open conifer forests with tall shrubs.  The model was able to distinguish hardwood trees, tall shrubs, shrubs, and conifers.  Figure 4 shows the classification result of a mix conifer, deciduous trees and short and tall shrubs near the lakeshore, classes that appear to have been accurately classified.

Figure 2: Classification results for regions of open conifer forests with shrub ground cover.

Figure 3: Classification results for regions of open conifer forests with shrub ground cover, as well as tall shrubs and hardwood trees.

Figure 4: Classification results for a heterogeneous region containing conifer and hardwood forests along the lakeshore.

Of the 1800 predictor variables found within a 15×15 window with 8 spectral bands, 122 variables were selected for use in the final model.  Of these 122 variables, five of the eight WV-2 spectral bands were found to be important: blue (400-450nm), red (630-690nm), “red edge” (705-745nm), and both near infrared bands (770-895nm, and 860-1040nm) (Figure 5).  Relative to the center pixel in a window, pixels used in the classifier were found across all distances-from-center, although there was a notable decrease in the number of variables chosen at distances between 2 and 5 pixels, with the most important radii being close to the center, and at the maximum distance from the center.  Figure 6 shows a histogram of variables chosen for a given range of distances from center vs. the total possible of variables possible at that distance. Figure 7 shows a histogram of the variables used as a function of azimuth. The distribution of the selected important variables in azimuth showed that most variables were picked up in the northwest and southeast directions. Figure 8 visualizes the spatial and spectral distribution of the selected variables used in the model.

Figure 5: Count of selected variables used by spectral band.  Band designations 1-8 represent respectively the Worldview-2 “coastal” (narrowband blue), blue, green, yellow, red, red edge, near-infrared 1 and near-infrared-2 spectral regions.

Figure 6: Relative importance of selected variables as a function of distance from center of the window.

Figure 7: Relative importance of selected variables as a function of azimuth from center of the window.

Figure 8: Visualization of the spatial and spectral distribution of the selected variables used in the model for the important spectral bands.

4. Discussion and Conclusions

In this study, we developed a generalizable spatialized classification approach to incorporate spatial information combined with spectral information together into a machine learning classifier.  The algorithm appeared to perform well even given the complex spatial and spectral patterns of the target classes.  Across all classes tested, the spatialized approach outperformed the classic spectral-only approach.  The biggest improvements over the spectral-only approaches were in the spatially complex vegetation classes (trees and shrubs).

The results of the variable analysis pointed towards a sensible choice of important variables. The distribution of the selected azimuths showed that most variables were picked up at northwest and southeast, which was along the solar azimuth at the time of the image was acquired. This makes sense, in that the model picked up the spatialized information such as shadows as a clue for the determination of a class. Tall conifer and hardwood trees have significant within-canopy shadows, whereas shorter shrubs, herbs, and non-vegetation classes do not.  The distance-from-center of the neighboring pixels chosen for use in the model also made sense in light of the size of the trees used in the training dataset. Conifer tree crown diameters varied from 2 to 10 meters in our study area, with an average of approximately 5m, or approximately 11 pixels, which was where the optimal window size appeared to be (Figure 1).  The presence of a large number of important variables at a large distance from center may indicate that the model is predicting based on the tree crown edges.  Generally speaking, the optimal search window size was greater than the size of the classification objects/classes.

The spatialized classification approach appears to overcome several challenges in classifying hyperspatial imagery. By taking spatial information into the classification scheme, the spectral value is not the only parameter in determining a class of a pixel, which overcame the high intra- and inter-class spectral variation problem. Shadows reported by previous studies (e.g. (Immitzer, Atzberger, & Koukal, 2012) as causing problems in classifying hyperspatial imagery not only did not cause errors in the analysis, but were leveraged with the spatialized feature extraction method, allowing the tree shadows to be valuable clues for the spatialized classification.  Conifer trees, for instance, were often confused with many of the other vegetation classes when only leveraging the pixel’s spectral data but was only somewhat confused with hardwood trees when using spatial features.  In general, the spatially complex classes (trees and tall shrubs) were rarely confused with the spatially simple, subpixel classes (short shrubs and herbs) when using spatialized classification.  We also not that problems of the border effects when using texture layers computed by box-shape moving windows did not occur in this proposed method even though the box-shape moving window was employed in extracting features.

Spatialized classification methods can be easily adapted to different landscape types and class types. The quality of training data was the most crucial part of the proposed approach. We found that it was important to capture a wide variety of spatial contexts for each of the classes (e.g. trees in open and closed canopies, different species of trees, etc.), and that multiple iterations with training data and visualizing results, refining and correcting errors. However, our results point towards a potential difficulty in porting a model from one image scene to another, particularly if the images are captured with different solar azimuths.  The development of azimuth-agnostic training data may be necessary to allow the production of a model from one scene’s data to be applied to another scene.

There are several aspects that could be used to improve classification performance. First, in our moving window approach, different shapes of the moving window such as circles or lines may be more appropriate than a square, which can bias the results towards diagonal distances. Second, while Random Forest was used in this study, the proposed framework should be somewhat agnostic towards the particular machine learning algorithm used, so other machine learning algorithms such as Support Vector Machine (SVM) or Neural Networks may lead to improved classification accuracies or computational efficiency. Finally, we suspect that incorporating spectral (e.g. NDVI) and spatial indices (distant pixel reflectance divided by center pixel reflectance) may improve model accuracy, so we intend to incorporate these into future models.

In conclusion, we demonstrated that the spatialized classification method could be leveraged to provide a generalizable and flexible approach to mapping vegetation in diverse ecosystems such as those found in the Lake Tahoe Basin. The output classification maps are able to begin to fill the information gap between large-scale Land Use/Land Cover type maps and detailed ground-based field inventories. We believe these approaches can provide improved and more detailed classification maps than are possible to derive from medium resolution remotely sensed imagery.  As the spatial resolution of imagery from newer earth-observing sensors continues to increase, many advanced techniques developed from fields like computer vision become more relevant to the remote sensing community. There is tremendous potential for interdisciplinary research projects to collaborate researchers in both fields to explore exciting techniques and raise interesting research questions and applications.


This research was performed, in part, as the core of a Master’s Thesis at the University of Illinois at Urbana-Champaign (UIUC).  Funding was provided by the George Beatty Fellowship (UIUC).  We thank Carlos Ramirez and Nathan Amboy at the USDA Forest Service R5 Remote Sensing Laboratory for providing image data and field crews, and to Susan Ustin and Paul Haverkamp at UC Davis for providing field crews as well as field equipment.  The authors would like to acknowledge the support of Research & Innovation and the Office of Information Technology at the University of Nevada, Reno for computing time on the Pronghorn High-Performance Computing Cluster.


Barbour, M. G., Keeler-Wolf, T., & Schoenherr, A. (2007). Terrestrial Vegetation of California. Brittonia (Vol. 31). doi:10.2307/2806144

Benediktsson, J. a, Swain, P. H., & Ersoy, O. K. (1990). Neural Network Approaches Versus Statistical Methods In Classification Of Multisource Remote Sensing Data. IEEE Transactions on Geoscience and Remote Sensing, 28(4), 540–552. doi:10.1109/TGRS.1990.572944

Bonan, G. B., Levis, S., Kergoat, L., & Oleson, K. W. (2002). Landscapes as patches of plant functional types: An integrating concept for climate and ecosystem models. Global Biogeochemical Cycles, 16(2), 5–1–5–23. doi:10.1029/2000GB001360

Breiman, L. (2001). Random Forests. Machine Learning, 45, 5–32.

Falkowski, M. J., Wulder, M. A., White, J. C., & Gillis, M. D. (2009). Supporting large-area, sample-based forest inventories with very high spatial resolution satellite imagery. Progress in Physical Geography, 33(3), 403–423. doi:10.1177/0309133309342643

Gougeon, F. A., & Leckie, D. G. (2006). The Individual Tree Crown Approach Applied to Ikonos Images of a Coniferous Plantation Area, 5, 1287–1297.

Greenberg, J. A., Dobrowski, S. Z., Ramirez, C. M., Tuil, J. L., & Ustin, S. L. (2006). A Bottom-up Approach to Vegetation Mapping of the Lake Tahoe Basin Using Hyperspatial Image Analysis. Photogrammetric Engineering & Remote Sensing, 72(5), 581–589.

Greenberg, J. A., Dobrowski, S. Z., & Ustin, S. L. (2005). Shadow allometry: Estimating tree structural parameters using hyperspatial image analysis. Remote Sensing of Environment, 97(1), 15–25. doi:10.1016/j.rse.2005.02.015

Hamada, Y., Stow, D. a., & Roberts, D. a. (2011). Estimating life-form cover fractions in California sage scrub communities using multispectral remote sensing. Remote Sensing of Environment, 115(12), 3056–3068. doi:10.1016/j.rse.2011.06.008

Immitzer, M., Atzberger, C., & Koukal, T. (2012). Tree species classification with Random forest using very high spatial resolution 8-band worldView-2 satellite data. Remote Sensing, 4(9), 2661–2693. doi:10.3390/rs4092661

Ishwaran, H., Kogalur, U. B., Blackstone, E. H., & Lauer, M. S. (2008). Random survival forests. Annals of Applied Statistics, 2(3), 841–860. doi:10.1214/08-AOAS169

Ishwaran, H., Kogalur, U. B., Gorodeski, E. Z., Minn, A. J., & Lauer, M. S. (2010). High-Dimensional Variable Selection for Survival Data. Journal of the American Statistical Association, 105(489), 205–217. doi:10.1198/jasa.2009.tm08622

Jin, S., Yang, L., Danielson, P., Homer, C., Fry, J., & Xian, G. (2013). A comprehensive change detection method for updating the National Land Cover Database to circa 2011. Remote Sensing of Environment, 132, 159–175. doi:10.1016/j.rse.2013.01.012

Key, T. (2001). A Comparison of Multispectral and Multitemporal Information in High Spatial Resolution Imagery for Classification of Individual Tree Species in a Temperate Hardwood Forest. Remote Sensing of Environment, 75(1), 100–112. doi:10.1016/S0034-4257(00)00159-0

Khoshgoftaar, T. M., Golawala, M., & Hulse, J. Van. (2007). An Empirical Study of Learning from Imbalanced Data Using Random Forest. 19th IEEE International Conference on Tools with Artificial Intelligence(ICTAI 2007), 2, 310–317. doi:10.1109/ICTAI.2007.46

Kim, M., Madden, M., & Warner, T. T. a. (2009). Forest Type Mapping using Object-specific Texture Measures from Multispectral Ikonos Imagery : Segmentation Quality and Image Classification Issues. Photogrammetric Engineering & Remote Sensing, 75(7), 819–829. Retrieved from

Lu, D., & Weng, Q. (2007). A survey of image classification methods and techniques for improving classification performance. International Journal of Remote Sensing, 28(5), 823–870. doi:10.1080/01431160600746456

Marceau, D. J. D. J., Howarth, P. J. P. J., Dubois, J. M. J. M., & Gratton, D. J. D. J. (1990). Evaluation Of The Grey-level Co-occurrence Matrix Method For Land-cover Classification Using Spot Imagery. IEEE Transactions on Geoscience and Remote Sensing, 28(4), 513–519. doi:10.1109/TGRS.1990.572937

Melgani, F., & Bruzzone, L. (2004). Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 42(8), 1778–1790. doi:10.1109/TGRS.2004.831865

Myint, S. W., Gober, P., Brazel, A., Grossman-Clarke, S., & Weng, Q. (2011). Per-pixel vs. object-based classification of urban land cover extraction using high spatial resolution imagery. Remote Sensing of Environment, 115(5), 1145–1161. doi:10.1016/j.rse.2010.12.017

Pu, R., & Landry, S. (2012). A comparative analysis of high spatial resolution IKONOS and WorldView-2 imagery for mapping urban tree species. Remote Sensing of Environment, 124, 516–533. doi:10.1016/j.rse.2012.06.011

Puissant, A., J. Hirsch, and C. Weber. (2005). The utility of texture analysis to improve per‐pixel classification for high to very high spatial resolution imagery. International Journal of Remote Sensing 26.4 (2005): 733-745.

Valero, S., Chanussot, J., Benediktsson, J. A., Talbot, H., & Waske, B. (2010). Advanced directional mathematical morphology for the detection of the road network in very high resolution remote sensing images. Pattern Recognition Letters, 31(10), 1120–1127. doi:10.1016/j.patrec.2009.12.018

Woodcock, C. E., & Strahler, A. H. (1987). The factor of scale in remote sensing. Remote Sensing of Environment, 21(3), 311–332. doi:10.1016/0034-4257(87)90015-0

Yu, Q., Gong, P., Clinton, N., Biging, G., Kelly, M., & Schirokauer, D. (2006). Object-based Detailed Vegetation Classification with Airborne High Spatial Resolution Remote Sensing Imagery. Photogrammetric Engineering & Remote Sensing, 72(7), 799–811. doi:10.14358/PERS.72.7.799

Zhang, Y. (2002). Problems in the fusion of commercial high-resolution satellite as well as Landsat 7 images and initial solutions. International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences, 34(4), 587-592.

With Our Resume Writing Help, You Will Land Your Dream Job
Resume Writing Service, Resume101
Trust your assignments to an essay writing service with the fastest delivery time and fully original content.
Essay Writing Service, EssayPro
Nowadays, the PaperHelp website is a place where you can easily find fast and effective solutions to virtually all academic needs
Universal Writing Solution, PaperHelp
Professional Custom
Professional Custom Essay Writing Services
In need of qualified essay help online or professional assistance with your research paper?
Browsing the web for a reliable custom writing service to give you a hand with college assignment?
Out of time and require quick and moreover effective support with your term paper or dissertation?