Landsat 8 six spectral band data and MODIS NDVI data for assessing the optimal regression tree models
Dates
Publication Date
2016-10-18
Start Date
2013-05-20
End Date
2013-08-28
Citation
Gu, Yingxin, Wylie, B.K., and Boyte, S.P., 2016, Landsat 8 six spectral band data and MODIS NDVI data for assessing the optimal regression tree models: U.S. Geological Survey data release, https://dx.doi.org/10.5066/F7319T1P.
Summary
In this study, we developed a method that identifies an optimal sample data usage strategy and rule numbers that minimize over- and underfitting effects in regression tree mapping models. A LANDFIRE tile (r04c03, located mainly in northeastern Nevada), which is a composite of multiple Landsat 8 scenes for a target date, was selected for the study. To minimize any cloud and bad detection effects in the original Landsat 8 data, the compositing approach used cosine-similarity-combined pixels from multiple observations based on data quality and temporal proximity to a target date. Julian date 212, which yielded relatively low "no data and/or cloudy” pixels, was used as the target date with Landsat 8 observations from days 140–240 in 2013. [...]
Summary
In this study, we developed a method that identifies an optimal sample data usage strategy and rule numbers that minimize over- and underfitting effects in regression tree mapping models. A LANDFIRE tile (r04c03, located mainly in northeastern Nevada), which is a composite of multiple Landsat 8 scenes for a target date, was selected for the study. To minimize any cloud and bad detection effects in the original Landsat 8 data, the compositing approach used cosine-similarity-combined pixels from multiple observations based on data quality and temporal proximity to a target date. Julian date 212, which yielded relatively low "no data and/or cloudy” pixels, was used as the target date with Landsat 8 observations from days 140–240 in 2013. The 30-m Landsat 8 composited data were then upscaled to 250 m using a spatial averaging method. Six Landsat 8 spectral bands (bands 1–6) at 250-m resolution were used as independent variables for developing the piecewise regression-tree models to predict the 250-m eMODIS NDVI (dependent variable). Furthermore, to ensure the high quality of the derived 250-m Landsat 8 data, and avoid any additional cloud and atmospheric effects, the percentage of 30-m pixels with “0” values within a 250-m pixel was calculated. Only those 250-m pixels with 0% of “0” values (i.e., all the 30-m pixels within a 250-m pixel have no zero values pixels) were selected to develop the regression-tree model.The 7-day maximum value composites of 250-m MODIS NDVI for the year 2013 were obtained from the USGS expedited MODIS (eMODIS) data archive (https://lta.cr.usgs.gov/emodis). Pixels with bad quality, negative values, clouds, snow cover, and low view angles were filtered out based on the MODIS quality assurance data to ensure high quality eMODIS NDVI data. The 2013 weekly NDVI data were then stacked and temporally smoothed using a weighted least-squares approach to reduce additional atmospheric noise. Temporal smoothing helps to ensure reliable NDVI values for weekly or longer cloudy periods. Finally, the 250-m MODIS NDVI for the date around Julian date 212 (similar date as the Landsat scene date) was extracted and used as the dependent variable in the regression tree model. ArcGIS, MS Excel, and Cubist are required to read and process the data.
Click on title to download individual files attached to this item.
RSD_meta_FGDC_f3.xml Original FGDC Metadata
View
15.29 KB
application/fgdc+xml
RSD.csv
506.29 KB
text/csv
Purpose
The data was used as an example to build the data-driven regression tree models and demonstrate the optimal data usage strategy that minimizes possible over- and underfitting effects in the regression tree models.
