Abstract

We investigated normalized difference vegetation index data from the NOAA series of Advanced Very High Resolution Radiometers and found several regions in United States that experienced anomalies over the course of the past 25 years. Our tasks consist of investigating and understanding the drivers of the anomalies in the Midwest forests of Minnesota and Wisconsin.

1. Introduction

Forests are known to absorb and sequester carbon dioxide of the atmosphere through the photosynthesis process. Globally forests contain 90% of the total vegetative carbon and 80% of the soil carbon and they assimilate 67% of the total terrestrial carbon dioxide (CO2) sequestered from the atmosphere (Landsberg and Gower, 1997). Hence this process is very essential in mitigating the global climate change. However this carbon is release into the air once the trees die from disease or cut down. The Advanced Very High Resolution Radiometer (AVHRR) Normalized Difference Vegetation Index (NDVI) is indicating a continuous disturbance in the forested area of Wisconsin and Minnesota (Neigh et al., 2008). Although a number of recent studies have found marked variations in NDVI throughout the Northern Hemisphere, they have not attributed these changes to regional factors that may include natural disturbances and/or human alterations to ecosystem functioning ([Gong and Shi, 2003], [Slayback et al., 2003], [Tucker et al., 2001] and [Zhou et al., 2001]). It is important to identify and quantify land cover type, because changes in land cover can alter ecosystem functioning and carbon storage ([Baldocchi and Amthor, 2001] and [Olson, 1975]).
NDVI (Tucker, 1979) is calculated from channel 1 (0.58–0.68 μm) and channel 2 (0.72–1.10 μm) from the NOAA AVHRR series of polar orbiting satellites as:

NDVI has been found to have a strong linear relationship to the fraction of photosynthetically active radiation (FPAR), the radiation that drives photosynthesis (0.4–0.7 μm) ([Myneni et al., 1995] and [Sellers, 1985]). FPAR is the main determinant of net primary productivity (NPP) of the ecosystem (Monteith, 1981). A number of factors are known to have negatively influenced the North American Midwest vegetation for example insects and pathogens outbreak (Radeloff et al., 1999) and severe drought ([Mo and Schemm, 2007]; [Sheffield et al., 2004]) leading to the decrease of rain fed agriculture. The purpose of this study is to understand and identify these causes of the anomalies in the Minnesota and Wisconsin forest.

2. Methods and Data * Study Area

According to the NDVI anomalies index generated in (Neigh, 2008), a large and dense trend of negative NDVI ranging from – 0.05 to less than – 0.1, is located in the neighboring forests in Central East Minnesota and Northwest Wisconsin. (Figure 1) * NDVI data

We used the bimonthly 8-km NDVI dataset processed by the Global Inventory Modeling and Mapping studies group (GIMMS). These data are produced from daily global 4-Km satellite data processed using the methodology described in Tucker et al. (2005) from NOAA’s AVHRR. Daily daytime data have been collected from 5 NOAA satellites to construct a time series of maximum bimonthly NDVI data spanning the period of January 1982 to 2007 for Wisconsin and Minnesota (Figure 2). The NDVI derived from the NOAA AVHRR instruments is one of the few global time-series satellite datasets with a 22+ year record to observe variability in climate through measurements of photosynthetic capacity (Schultz and Halpert 1995, Suzuki et al. 2000, Lotsch et al. 2003).

Figure 1: Minnesota and Wisconsin NDVI, Landsat Path-Row, and Meteorological stations Map

A number of non-target conditions can influence AVHRR NDVI. Factors include: sensor degradation; atmospheric attenuation; and solar zenith angle effects from orbital ‘drift’ (Neigh, 2007).
The NDVI data was then processed through a time series plot in order for us to understand the impacts of the anomalies each year. This helped in selected to proper Landsat images to validate NDVI visibility through a 30-km radiometer (Neigh et al. 2007).

* Meteorological Data

Daily meteorological data for mean surface air temperature and precipitation were analyzed for 14 stations (Figure 1) from the United States Historical Climatology Network (USHCN) for the 1982 to 2005 time period (http://cdiac.ornl.gov/epubs/ndp/ushcn/ushcn.html). The meteorological data was used to validate the drought periods of 1987 – 1989, 2000 – 2003, and 2005 – 2006 (Mo and Schemm, 2007, Sheffield et al., 2004). Thus the NDVI anomalies were correlated to the climate data obtained through a chart algorithm explained in Neigh et al., 2008. * Agricultural Data
The USDA National Agricultural Statistics Services (NASS) crop data were acquired in order to determine whether the crop productivity was affected by the drought in the late 1980s and early 2000s. Our hypothesis is that the widespread drought periods produced a negative anomaly trend and reduced rain fed agriculture production limiting area planted and harvested. (Andreadis et al., 2005, Sheffield et al., 2004) * Insects Outbreak Data
Several studies have been done on the impacts of insects on the forested part of the study area [(Radeloff et al., 1999), (Radeloff et al., 2000), deBeurs & Towsend, 2008]. A lot more articles dealt with monitoring the progress and the aftermath of the insect defoliation. [(Coops et al, 2006), (Radeloff et al. 2003), and Wolter et al. 2008)].
The analysis consisted of first obtaining forest disturbance data related to insects defoliation from the Forest Inventory and Analysis (FIA) division of the United States Department of Agriculture (USDA) Forest Services office (http://fia.fs.fed.us/). Then map layers should be generated to spatially identify the extent and impact of the defoliation on the trees along with the defoliator insects. * Landsat Data

Landsat data were acquired for the same areas that had NDVI anomalies that were less than 0. Landsat Multispectral Scanner (MSS) Global Land Survey (GLS), Thematic Mapper (TM) AND Enhanced Thematic Mapper (ETM) images since the late 1970s were downloaded from the United States Geological Survey (USGS) Glovis server (www.glovis.usgs.gov). Late spring and summer were the seasons chosen because the vegetation is grown at that time of the year, therefore there would be detectable by the landsat sensors. We mainly focused the path row 27_28 and 26_28 because the anomalies are really dense in these areas (Table 1)
Neigh et al, 2008 developed a methodology to determine land cover for three time periods to quantify land cover changes that could be responsible for trends in NDVI. MSS/GLS/TM Change Detection images | | | | | Path/Row | Map Date (Max Cloud = 10%) | Cloud % | MSS/GLS/TM | | 9-May-1975 | 10 | MSS | 29/28 | 6-Sep-1976 | 0 | MSS | | 2-May-1978 | 0 | MSS | | 1-Sep-1974 | - | GLS | | 19-Jul-1975 | 10 | MSS | | 6-Aug-1975 | 0 | MSS | 28/28 | 11-May-1976 | 0 | MSS | | 23-Sep-1976 | 10 | MSS | | 10-May-1978 | 10 | MSS | | 17-Sep-1979 | - | GLS | 27_28 | 9-May-1989 | - | TM | | 6-Sep-1992 | - | TM | 26_28 | 29-Apr-1988 | - | TM | | 28-Aug-1991 | - | TM | Table 1: Landsat MSS/GLS/TM Chart
When using multi-spectral data, changes in surface reflectance have been associated with changes in vegetation cover and extent. Generally, a higher reflectance is associated with sparse vegetation cover, and a lower reflectance is associated with dense vegetation or water in visible wavelengths. When investigating pixel response to changes in land cover, deforestation will typically increase pixel brightness (darker vegetation to lighter soil) whereas afforestation and succession would decrease pixel brightness (bare soil to vegetation). (Jensen, 2006). We use this simple observation to quantify variance in land cover. Radiance values are used in our analysis because the selected study sites have marked changes in land surface red and near infrared spectral reflectance observed by coarse resolution 8 km AVHRR. The spectral changes observed between vegetated and non-vegetated pixels at Landsat 30 m resolution visible with 8 km AVHRR far outweigh the influences of sun angle, variability in atmospheric attenuation, and sensor degradation.
The Tassel Cap (TC) change detection method has been applied to Landsat data acquired.
These method identify change in multi-temporal data and have been enhanced with hybrid methods ([Guild et al., 2004], [Jin and Sader, 2005], [Lanjeri et al., 2004], [Lunetta et al., 2002], [Nackerts et al., 2005], [Rigina, 2003] and [Warner, 2005]). To adequately account for temporal changes in vegetation cover our analysis developed two methods to stratify multi-spectral observations into thematic maps of land cover and land cover change. The first method was developed for the boreal zone where variance in red and near-infrared spectral reflectance is observed from forest to non-forest changes. Method two required adaptation to identify interannual land cover land use changes that may be difficult to distinguish in regions of intense agriculture.. Both methods are based on change detection algorithms currently in use ([Guild et al., 2004] and [Lanjeri et al., 2004]).
A base map was first generated from the 1970 image from the red (Channel 3), Green (Channel 2), and Near infrared (Channel 4) from the Landsat MSS. An unsupervised ISODATA classification was performed on these channels due to their ability to discriminate vegetation density and type while exhibiting reduced atmospheric contamination that commonly affects the blue and green visible wavelengths. ISODATA is a standard clustering algorithm available in most image processing software packages and is based on procedures in which cluster centers are iteratively determined sampled means (Tou & Gonzales, 1974). If a scene contains atmospheric constituents, multiple iterations of the classification were performed to mask and eliminate cloud, cloud shadow, and haze cover. This was necessary to minimize atmospheric contamination that alters the class distribution structure. That was the basis for selecting and distinguishing the classification cover types. Once a base map of recent land cover was derived, we reverted in time to define locations of change in land cover and mask the current thematic map for prior land cover types. A similar method has been performed by Lanjeri et al. (2004). This method was applied to reduce misclassifications between dates with no land cover changes. Change detection was subsequently performed once a base thematic map had been developed for each anomaly area. A linear tassel cap transformation was first applied to the multi-date images reducing multi-spectral redundancy to indices of brightness, greenness, and wetness ([Crist and Cicone, 1984] and [Huang et al., 2002]). The linear tassel cap also enhanced differences in brightness, greenness, and wetness that may occur between multi-date images. The weights of the linear tassel cap transformation were fixed, were sensor specific, and were not scene dependent (Guild et al., 2004). The sensor specific weights of the linear tassel cap transformation aid in normalizing between sensors for change detection analysis. Finally, an unsupervised ISODATA clustering algorithm was performed to all transformed images to group similar spectral vectors into ‘change’ and ‘no change’ clusters from the bi-temporal images (Richards, 1993). The same procedure was also performed in agriculture regions with the modification of adding an additional pair of images from the same year as the base period of investigation (Figure 3). This was done to capture crop rotation during a growing season while distinguishing irrigated agriculture from fallow croplands and natural grasslands. Irrigated agriculture generally had an enhanced signature of wetness and greenness compared to non-irrigated vegetation in the semi-arid high plains environment. Appendix 3 shows the change detection output. The output generated does not display any major land cover change that could affect NDVI negatively.

Figure 3: Land cover change classification algorithm diagramming steps 3. Results and Discussions * NDVI
NDVI anomalies time series in the AVHRR data generated (Figure 4) shows a disturbed trend in the year 1990 to 1995 but also in the late 1999 to 2005. * Meteorological Data The meteorological dataset output in a time series plot comparing the temperature and the precipitation against NDVI did not show any correlation in respect of NDVI disturbance. (figure 5) Figure 4: NDVI Time Series. * Agricultural Data
The NASS crop analysis for Wisconsin and Minnesota did not show any trend with NDVI due to lack of needed data. * Insect Outbreak
After failure to receive any NDVI trend in respect to meteorological data and agricultural data, the investigation of the insect defoliation validate that insects have been a cause to tree mortality in Minnesota and Wisconsin. Appendix 1 displays the acreage of defoliation by major insect from 1995 to 2006. And Appendix 2 shows the time series of defoliation acreage by insect from 1995 to 2006.
From 2000 to 2002 a cumulative of 7 million acres was strongly affected by the Forest Tent Caterpillar. The Forest Tent Caterpillar is a indirect tree mortality agent. The main damage it does is weakening the tree defense system, which makes it vulnerable to disease and pathogens attacks. (Minnesota DNR, 2002)

Figure 5: Wisconsin Met. Stations Diagram the color red is for temperature while the blue is for precipitation. The NDVI trend is not disturbed according this chart output

4. Conclusion
No conclusion can be drawn at this stage of the project as we are still investigating the major drivers of NDVI anomalies in Wisconsin and Minnesota. 5. Future Tasks
As of now, we are currently working to validate the insect outbreak with spatial data by generating other disturbance index method used in (Hais, 2009). The new disturbance index will then be compared to the NDVI in order to identify a similarity between the anomaly zones.