Researchers Track A Puzzling Pattern Of Tree-killing Insects

This image shows the location of the LIDAR study in the Ozark National Forest near White Rock Mountain. The researchers hope to determine patterns in the red oak borer infestation that may lead to new forest management practices to deal with the beetles.

Credit: Jason Tullis


The red oak borer, a middle-sized, nocturnal brown "long-horned" beetle, lived in relative obscurity in red oak trees in the Ozark Mountains until 1999, when forestry professionals and researchers began noticing oaks dying in droves. When Fred Stephen, University Professor of entomology, and his students began examining the trees, they found them filled with borers.

gement practices.
The red oak borer, a middle-sized, nocturnal brown "long-horned" beetle, lived in relative obscurity in red oak trees in the Ozark Mountains until 1999, when forestry professionals and researchers began noticing oaks dying in droves. When Fred Stephen, University Professor of entomology, and his students began examining the trees, they found them filled with borers.

The red oak borers have a two-year life cycle, spent mostly as larvae that bore into the heartwood of the host oaks. The larvae carve out galleries in the wood, chewing through layers of rings in the middle of the tree and creating small holes.

Most of the time natural controls on population growth, including the defenses that oaks mount, successfully combat the larvae. But an unexplained dramatic increase in larval density -- from an average three or four to a tree to 70 or 80 in a tree -- led to the deaths of tens of thousands of trees.

Then in 2005, the red oak borer population dropped 98 percent, returning to typical levels.

"We're tying to understand why all of this happened," Stephen said.

To do so, Stephen and geosciences professor Jason Tullis have taken their research into the air. Tullis is working with light detection and ranging (LIDAR), which uses pulses of laser light to accurately map a forested area of the Ozark National Forest near White Rock Mountain, where many trees succumbed to red oak borers during the infestation.


The LIDAR instrument, mounted in an airplane, shoots about 50,000 near-infrared laser pulses per second at the earth. Depending on the path of a given laser pulse, up to four return pulses are recorded by the instrument's receiver. In the case of a typical red oak tree, return pulses originate from leaves, branches, the trunk and the ground.


Using GPS, gyroscopes and other instrumentation, the location where each LIDAR return pulse originated is computed, allowing the researchers to study a three-dimensional "point cloud" representing forest structure and terrain. This information will be used in conjunction with field studies to look for patterns that might provide insight into the origins of the outbreak.

"If we could map the vulnerability of these ecosystems, we might be able to determine what areas need extra attention," Stephen said. Practices such as thinning the forest or changing species composition might help create forests less vulnerable to such infestations.

"GIS and remote sensing can help find a pattern, and this can help us understand the underlying science of the ecosystem," Tullis said.



Story Source:

The above post is reprinted from materials provided by University Of Arkansas. Note: Materials may be edited for content and length.

Article source: Science daily

Wireless sensor network monitors microclimate in the forest

A wireless sensor network records data such as soil and leaf humidity, as well as air temperature.

Credit: Image courtesy of Fraunhofer-Gesellschaft

During a forest monitoring operation, forestry scientists measure various environmental values. This is how they obtain indications about how the forests are changing and what can be done to preserve them. However, installing and maintaining the wired measuring stations is complex: Researchers developed a wireless alternative.


What effect does climate change have on our local forests? What types of trees will be suitable for which geographic location? And how great is the pollution level here? Forestry scientists are conducting "forest monitoring" procedures: They continuously record parameters such as soil humidity or pollutant penetration at permanently installed monitoring stations. The results of such examinations contribute to maintaining the ecological stability of the forests over the long term. The problem: Not only are the wired measuring devices complex to install and maintain, they also hinder silvicultural work in the forest.

In the future, technologies from the Fraunhofer Institute for Microelectronic Circuits and Systems IMS could enable differentiated analysis without any bothersome cables. Scientist from the institute in Duisburg installed a new type of system for microclimatic monitoring on the grounds of the Northwest German Forestry Testing Facility in Göttingen, Germany. "We are using a wireless sensor network so we can measure relevant parameters within an area at many sites simultaneously," explains Hans-Christian Müller, group manager at the IMS. This way, we receive a very detailed picture about the environmental conditions on site, without much installation effort. Depending on which values they are to measure -- for example, soil moisture content, air temperature or the moisture in the leaves -- different sensor nodes are inserted into the soil or affixed to branches. If required, the measuring positions can be changed without much effort. The intelligent mini-computers automatically form a network and control the transmission of measurement data within this network. The results are transmitted by cellular radio to a central tree stock database. To facilitate this, a mobile cellular modem is connected directly to the sensor network.

Providing power to the sensor nodes poses a particular challenge. Mounting solar cells to the sensors -- a favored solution in other agrarian and forestry applications -- is not an option due to the low penetration of sunlight under the leafy canopy of the trees. That's why, to date, there has been no alternative to batteries that have to be replaced regularly. Researchers, however, managed to significantly increase battery life, keeping maintenance requirements within reasonable limits: "We adapted the software design accordingly and now have operating times of 12 months," says Müller. A software solution integrated into the sensors ensures that the radio nodes are for the most part in an energy-saving sleep mode. They are active only during the measurement and data transmission process. The measurement intervals can be set to be variable. Parameters that change slowly such assoil moisture need not be measured as often as air temperature, for example, which is subject to larger variations. Since data transmission requires the most energy, the measurement values are calculated as early as the sensor node. This reduces the data volume.

The new technology is already in use in Göttingen as part of the joint project "Smart Forest." The project aims to optimize forestry processes with the aid of microelectronic components. The researchers from IZM will be introducing their results on the "Smart Forests" as well as other developments on the industrial application of wireless sensor networks at the Messe Sensor + Test tradeshow from June 7 -- 9 in Nuremberg, Germany.



Story Source:

The above post is reprinted from materials provided by Fraunhofer-Gesellschaft. Note: Materials may be edited for content and length.

Article source: Science Daily

Forest Canopy Density Classification Using Texture Quantization of Panchromatic Aerial Images

By Hamed Ashoori, MasoudTaefi Feijani, Valadan Zoej

ABSTRACT

Forest canopy density is an important criterion for forestry application. Several methods were introduced to compute these criteria. None manual methods use multispectral images to determine the canopy density. It also could be extracted from Arial images manually. Arial images are valuable data with rather high special resolution, but some of them are panchromatic and are not suitable for spectral processing. Image texture which is valuable contextual information helps the interpreter to distinguish different canopy density areas. In this paper panchromatic Arial images were used to classify forest canopy density cover. Texture features were generated from image and use beside the image in classification. The results shows good improvement in classifying different canopy density covers.

INTRODUCTION

The importance of studying,Forest is a very complex ecosystem. The complexity of forested area in Iran is more than similar area in the world. The Caspian Hyrcanian mixed forest in the north of Iran has a very great diversity. UNESCO classifies the Forested area in Iran as a natural world heritage sites for their great age and diversity. Forest in Iran is habitat of many endemic and semi-endemic and relic species; these exclusive properties make difficult processing and analyzing of satellite images.

Different methods were introduced to estimate forest canopy density; some of them are based on multispectral images which are rather expensive for large areas. Arial panchromatic images have been captured approximately for all the parts of Iran. These images could not be used for spectral processing, but has valuable information which could be used for interpretation and classification of forest canopy density visually. Sotexture quantification could be used to generate new features from panchromatic image, then they are used with the source panchromatic band as input data for classification.

Forest Canopy Density Estimation Models
Since now many models have been used for estimation forest canopy density and biomass inventory from satellite images. Forest canopy density or FCD is a very important factor for forest management and assessment. Some of the general methods for this issue is explained in short term as below:

Artificial neural network (Boyd et al., 2002),Artificial neural networks are neurologically inspired statistical mechanisms also employed in classification of forest cover using various sensors (Boyd et al., 2002). A three layer feed forward error-back propagation artificial neural network implemented in Interactive Data Language (IDL) was used in order to predict forest canopy density as a continuous variable. The algorithm minimizes the root mean square error between the actual output of the multi-layered feed forward perceptron and the desired output (Skidmore et al., 1997). Following Atkinson and Tatnall (1997) to search for system parameters to increase the accuracy of the method and avoid overtraining of the neural network. The neural networks with the sub sample of 186 sites with canopy density and the seven ETM+ bands was trained and the best combination of optimum learning rate and momentum to minimize the root mean square error (RSME) was empirically established. The best results were obtained with a learning rate of 0.7, a momentum of 0.7 and four hidden nodes. The RSME stabilized after approximately 7000 epochs. Finally, 20 iterations of 7000 epochs were performed and the best one was selected based on root mean square error.

Multi Linear regression techniques (Iverson et al., 1989 ; Levesque and King, 2003),Multiple linear regression techniques have been used to model the relation between spectral response and closed canopy conifer forest cover (Ripple, 1994). In this study, a multiple linear regression model has been developed, which best described the relation between canopy density and the seven ETM+ spectral bands. The regression equation using n = 186 observations is:

Y= 3.32 + 0.021B1 – 0.002 B2 + 0.003 B3+0.024 B4+0.023 B5+0.021 B6-0.029 B7 (1)
Where Y is the predicted forest canopy density and B1–B7 is the reflectance value of bands 1–7 of Landsat ETM+ image.

Forest canopy density mapper (Rikimaru, 1996), Rikimaru introduced an alternative deductive approach, i.e. forest canopy density mapper to map forest canopy density using four indices (vegetation, bare soil, shadow and surface temperature) derived from Landsat TM imagery.
Based on these four variables, nine canopy density classes, namely 0, 1–10, 11–20. . . 71–80+ were obtained. This model involves bio-spectral phenomenon modeling and analysis utilizing data derived from four indices, namely:

• advance vegetation index (AVI)
• bare soil index (BI)
• shadow index (SI)
• Thermal index (TI).

Using these four indices the canopy density for each pixel was calculated in percentage
The method requires intervention by an operator to establish threshold values. The accuracy obtained in three SE Asian countries averaged 92% (Rikimaru, 1996).
Maximum likelihood classification (MLC),as a parametric classifier, maximum likelihood classification method calculates the probability that a given pixel belongs to a specific class and assigns the pixel to the class having the highest probability (Richards, 1999). The training set of 186 pixels into 10 canopy classes, namely 0, 1–10, 11–20, . . ., 71–80+ was classified. The Interactive Data Language (IDL 6.0) and ENVI 4.0 (ENVI, 2003) was used for image classification. (Chudamani Joshi, 2005)
They are the more general models used for forest canopy classification. Some of others methodologies are:

• Object based classification (Dorren et al., 2003)
• Decision tree classification (Souza et al., 2003)
• Spectral unmixing at pixel or subpixel scale (Cross 1991)

General Models Used in Operational and Research Project in Iran
In international scale many models have been used in operational projects. Great center and organizations such as USGS, CRC, ITTO& FAO used one of the mentioned models for their activities under operational large projects.
In IRAN only research project that applied by universities or research institutes used professional models and algorithms.
Administrative agencies and executive organization used very elementary and basic models for their simplicity of implementation. Our study shows that 3 model (or indices) is very popular in these projects:

• NDVI
• Principle Component Analysis
• Visual Interpretation

Along with simplification of these models and indices the main reason for this issue is lack of satellite and field data in Iran. Indeed high spatial resolution data of natural and forested area is very low besides related lack of simultaneous field and train data.

Aerial photo interpretation of natural resource is a common activity that is being followed from near 1960 till now. And mid resolution images are accessible from 1973 (1 year after launching landsat#1).although any model that can use the aerial photo as a source of data not face with lack of image data, because as mentioned above they are taken from 1960 by 5-10 year of intervals. According to these capacities and limitations we use Panchromatic Arial Images in our model as aninput image data.

DATA SPECIFICATIONS
Two panchromatic Arial stereo images from Zagros MountainsGavbarg region inYASUJ province area where used. Images were captured at 1999 in 1:40,000 scale. At first images was oriented using ground control points, then DEM were generated in overlap area. Ortho image were generated using one oriented image and generated DEM.
Train and check data were selected using manually classified image. (Figure 1)
Sixclasseswere selected based on their canopy density which was determined manually, and 500*500 subsets were selected around each class.Selected classes were very dense canopy to bare land; they were ranked based on the canopy density. The test image was generated by mosaicking the 6 subsets. (Figure 2)

Feature generation
Different methods were introduced by authors for quantifying image texture. These methods could be used to generate image base features. Generated features could improve image classification accuracy beside spectral features.
Several methods were used to generate images, mean and variance from first statistical methods, direct variogram and madogram from geostatistical methods, low-pass and high-pass ringing and slice filters from Fourier based filters. These features use the following equations to generate features.

First Order Statistical Features
If (I) is the random variable representing the gray levels in the region of interest, the first order histogram P (I) is defined as (Theodoridis, 1999):

Now different features can be generated by using the following equations.

Moment

Where Ng = number of gray levels.
is the simple mean of pixels. Also 2nd, 3rd and other moments can be used.

Central Moments
(3)

Geostatistical Features
Geostatistics is the statistical methods developed for and applied to geographical data. These statistical methods are required because geographical data do not usually conform to the requirements of standard statistical procedures, due to spatial autocorrelation and other problems associated with spatial data (http://www.geo.ed.ac.uk).
Semivariogram that represents half of the expectation of the quadratic increments of pixel pair values at the specified distance can quantify both spatial and random correlation between the adjacent pixels. (Goodenough, et al, 2003) It is defined as:
(4)



That is the classical expression of variogram (h) here represents a vectorial lag between pixels. In this study direct variogram, madogramvariogram have been used.

Direct Variogram

In this approach the following equation is used to estimate:

(5)



n(h) is the number of pairs that are in mask filter.
Madogram

This is similar to direct variogram except squaring differences, but uses the absolute value of differences.
(6)

Fourier Based Features
Fourier transformation, transforms a signal from space/time domain to frequency domain. The amplitude and phase coefficients are two outputs of a Fourier transformation. So different texture patterns could be identified by their Fourier coefficients but because in this research one value for each pixel is required, raw Fourier coefficients couldn’t be used. Several features can be generated using sum of the Fourier amplitude under different masks (Pratt, 2001). These are comprised ringing, sectorial, horizontal and vertical which are shown in figure 2.

(7)



(8)



Figure3. Different mask which can be used to generate features from Fourier coefficients
Different parameters can be set in each method; the main parameter is window size. Different window sizes were used to generate features in each method.All features were generated using
3*3, 9*9, 15*15, 21*21, 27*27 and 33*33 window sizes. In geostatistical method four lags were used, they are [1, 0], [1, 1],[1,1] and [-1,1] also in Fourier method, two masks which are shown in figure 2 were used and high frequency and low frequency features were generated using each mask.

Implementation, Results and Conclusion
To evaluate the effect of using generated features in classification process, firstly the image was classified using dn slicing (parallel pipe classification) because the input is a gray-scale panchromatic image and couldn’t be classified using other methods.
All possible features using mentioned methods and selected window sizes and lag or mask parameter were generated. Then classification was done using each generated image beside panchromatic image as input data. Accuracy assessment was done, generating confusion matrix and computing overall, kappa and producer accuracies.
Table 1 shows parallelepiped results and best results obtained for each class user accuracy and overall and mean accuracy.

Results shows that using generated features could separate different forest canopy densities better and increase classification accuracy. It could be said that different methods for generating image based features should be used for different aims (e.g. Madogram for class 2); also some features could be used for general improvement (Mean with 33*33 window size).

References

References from Journals

Atkinson, P.M., Tatnall, A.R.L., 1997. Introduction neural networks in remote sensing. Int. J.Remote Sensing 18, 699–709

Boyd, D.S., Foody, G.M., Ripple, W.J., 2002. Evaluation of approaches for forest cover estimation in the Pacific Northwest, USA, using remote sensing. Appl. Geography 22, 375–392

Chudamani Joshi, Jan De Leeuw, Andrew K. Skidmore, Iris C. van Duren, Henk van Oosten, 2005, “Remotely sensed estimation of forest canopy density: A comparison of the performance of four methods”, International Journal of Applied Earth Observation and Geoinformation

Cross, A.M., Settle, J.J., Drake, N.A., Paivinen, R.T.M., 1991. Subpixelmeasurement of tropical forest cover using AVHRR data. Int. J. Remote Sensing 12, 1119–1129.

Haralick, R.M., Shanmugam, K., Dinstein, I., 1973. “Textural features for image classification”, IEEE Transactions on Systems, Man and Cybernetics, vol. 3, no. 6, pp 610-621.

Iverson, L.R., Cook, E.A., Graham, R.L., 1989. A technique forextrapolating and validating forest cover across large regions:calibrating AVHRR data with TM data. Int. J. Remote Sensing 10,1805–1812

Levesque, J., King, D.J., 2003. Spatial analysis of radiometric fractions from high-resolution multispectral imagery for modelling individual tree crown and forest canopy structure and health. Remote Sensing Environ. 84, 589–602.

Skidmore, A.K., Turner, B.J., Brinkhof, W., Knowle, E., 1997. Performance of a neural network: mapping forests using remotely sensed data. Photogrammetric Eng. Remote Sensing 63, 501–514.

Souza Jr., C., Firestone, C.L., Silva, L.M., Roberts, D., 2003. Mapping forest degradation in the Eastern Amazon from SPOT-4 through spectral mixture models. Remote Sensing Environ. 87, 494–506.

References from Books
John A. Richards, 1999, “Remote Sensing Digital Image Analysis an Introduction”, Springer-Verlag

Pratt, 2001,” Digital Image Processing”

SergiosTheodoridis, 1999, “Pattern Recognition”, Academic Press
References from Other Literature

Ashoori, H., Alimohammadi, A., ValadanZoej, M. J., Mojarradi, B., 2006. Generating Imagebased Features for Improving Classification Accuracy of High Resolution Images, May, ISPRS Mid-term Symposium, Netherlands.

Dorren, L.K., Maier, A.B., Seijmonsbergen, A.C., 2003. Improved Landsat-based forest mapping in steep mountainous terrain using object-based classification. Forest Ecol. Manage. 183, 31–46.

Goodenough, David, A.S. Bhogal, R. Fournier, R.J. Hall, J. Iisaka, D. Leckie, J.E. Luther, S.Magnussen, O. Niemann, and W.M. Strome, Earth Observation for Sustainable Development of Forests (EOSD), Victoria, B.C.: Natural Resources Canada, http://www.aft.pfc.forestry.ca,1998

P.S. Roy, S. Miyatake and A. Rikimaru, “Biophysical Spectral Response Modeling Approach for Forest Density Stratification”, ACRS 1997

Rikimaru, A., 1996. Landsat TM data processing guide for forest canopy density mapping and monitoring model. In: International Tropical Timber Organization (ITTO) workshop on utilization of remote sensing in site assessment and planning for rehabilitation of logged-over forest, Bangkok, Thailand, pp. 1–8.
References from websites

http://www.geo.ed.ac.uk

Acknowledgements
Visual classified image which have been used as training source is received from “Forests, Range and Watershed Management Organization (FRWO); Engineering and Evaluation Bureau”

Source: Coordinates 

Forests on diet – the map of global forest extension

By Stefan Mühlbauer



A new high resolution map of global forest extension covering the time span 2000 – 2012 was recently presented by Department of Geographical Sciences, University of Maryland, US. The time series is based on 654.178 LANDSAT images resulting in a global wide map displaying forest change at a never seen spatial detail of down to 30m. Thus, the map entails globally ‘consistent but locally relevant information’, according to a geographer of University of Maryland. Indeed, the map is useful for extracting information on local forest change, while potentially every corner of the globe may be entered. The huge amount of data processing was possible only through cloud computing.

Methodologically, forests were considered as all vegetation taller than 5m and are expressed as a percentage per output grid cell as ‘2000 Percent Tree Cover’. ‘Forest Loss’ is defined as a stand-replacement disturbance, or a change from a forest to non-forest state. ‘Forest Gain’ is defined as the inverse of loss, or a non-forest to forest change entirely within the study period.

The new forest map reveals that between 2000 and 2012 2.3 millions km² of forest havevanished. To present the gain and loss more clearly I am going to state the raw numbers:

2000 – 2012 global forest dynamics

• Gains: 800.000 km²
• Losses: 2.300.000 km²
• Loss and Re-gain: 200.000km²

The greatest amount of loss still occurred in the tropics that count for 32% of all losses. While in Brazil due to political efforts the rate of loss reduced slightely (though after 2012 the restrictions for deforestation were loosened up again), the deforestation rate in Indonesia doubled after 2003 from 10.000 km² to more than 20.000 km² forest cut per year. Considerable are also the losses in the Canadian and Russian boreal forests.

Forest monitoring belongs to one of the highly significant topics of today. Initiatives such as UN’s REDD+ highlight the need for information upon forest change and biomass. Forests impact the climate (CO2 household), biodiversity of plants and animals, but also the humans in a positive manner. A researcher of the mapping team found out that tree cover correlates with human health as people living close to forests eat a healthier diet than people in other environments do (FAO article1, FAO article2).

In an increased situation of urbanisation, loss of biodiversity and enhanced consumption of resources the protection of forests as ecological regulators is of great importance. As political desicions for stopping deforestations unfortuantely need hard facts those forest and biomass monitoring programs in my opinion are strongly necessary in order not to experience forests being on diet themselves!

The new global map of deforestation reveals 2.3 million square kilometers lost between 2000 and 2012. Red shows losses, blue gains, purple loss and gain.

Indonesia lost forests the fastest of any nation between 2000 and 2012. Red shows losses, blue gains, purple gain and losses.Credit: Image courtesy Matt Hansen, University of Maryland

Forest losses in tropical South America between 2000 and 2012. Particularly at the southern edge of the Amazonian Basin, in Bolivia, Paraguay the loss of forest are considerable. Red shows losses, blue gains, purple losses and gains.

A map of change in North American forests between 2000 and 2012. Red is loss and pink represents areas of loss and gain.

Credit: Image courtesy Matt Hansen, University of Maryland

Losses in the Canadian boreal forests in a more detailed view. Red shows losses, bllue gains, purple losses and gains

.

Forest change in Europe: A wind storm in 2009 leveled a forested area in the south-west of France. Portugal exhibits a strong dynamic of forest loss and gain. Red shows losses, blue gains, purple losses and gains.

Source: Geoawesomeness

Near-Real Time Delivery of MODIS-Based Information on Forest Disturbances

Journal articly by: Robert A. Chastain, Haans Fisk, James R. Ellenwood, Frank J. Sapio, Bonnie Ruefenacht, Mark V. Finco, Vernon Thomas





Abstract
The Real-Time Forest Disturbance (RTFD) program of the Forest Service, U.S. Department of Agriculture (USFS) provides timely spatial information regarding changes in forest conditions to the Forest Health Protection (FHP) and State and Private Forestry (S&PF) community for improving aerial detection and forest health survey efficiency. The USFS Remote Sensing Applications Center (RSAC) creates CONUS-wide forest change geospatial layers for the RTFD program every 8 days during the growing season using image data from the Moderate Resolution Imaging Spectroradiometer (MODIS), and delivers these data to a web mapping application named the Forest Disturbance Monitor (FDM) developed by the USFS Forest Health Technology Enterprise Team (FHTET).

Differences in the timing, duration, and severity of disturbances in forested landscapes result in a broad array of possible types of forest change. Two effective remote sensing change detection approaches using MODIS satellite data are employed to detect and track quick and ephemeral change as opposed to gradually occurring disturbances in forest health. The first uses a statistical (Z-score) change detection approach designed to discern intraseasonal ‘quick’ changes in forest conditions caused by events such as defoliations or storm damage. The second approach uses trend analysis to identify areas where slower, multiyear changes occur in forested areas, such as bark beetle outbreaks and drought stress in the western coniferous forest biome.

References

Adams HD, Macalady AK, Breshears DD, Allen CD, Stephenson NL, Saleska SR, Huxman TE, McDowell NG (2010) Climate-induced tree mortality: earth system consequences. EOS Trans Am Geophys Union 91(17):153–154CrossRef

Hargrove WW, Spruce JP, Gasser GE, Hoffman FM (2009) Toward a national early warning system for forest disturbances using remotely sensed canopy phenology. Photogr Eng Remote Sens 75(10):1150–1156

Kennedy RE, Cohen WB, Schroeder TA (2007) Trajectory-based change detection for automated characterization of forest disturbance dynamics. Remote Sens Environ 110:370–386CrossRef

Nielsen EM, Finco MV, Hinkley E (2008). Change detection in image time-series affected by directional reflectance and phenological variability: application to forest disturbance monitoring. In: Proceedings of the 2008 IEEE International Geosciences and Remote Sensing Symposium, Boston, 6–11 July 2008

Raffa KF et al (2008) Cross-scale drivers of natural disturbances prone to anthropogenic amplification: the dynamics of bark beetle eruptions. Bioscience 58(6):501–517CrossRef

Ruefenacht B, Finco MV, Nelson MD, Czaplewski R, Helmer EH, Blackard JA, Holden GR, Lister AJ, Salajanu D, Weyermann D, Winterberger K (2008) Conterminous U.S. and Alaska Forest Type Mapping Using Forest Inventory and Analysis Data. USDA Forest Service—Forest Inventory and Analysis (FIA) Program & Remote Sensing Applications Center (RSAC)

Vermote EF, Roy DP (2002) Land surface hot-spot observed by MODIS over Central Africa. Int J Remote Sens 23:2141–2143CrossRef

Source