Normalized Difference Water Index (NDWI)

Radiance from a satellite platform is strongly affected by the presence of the atmosphere. Atmospheric correction for NDWI was made to convert satellite-based radiance to an estimate of ground reflectance. Please see the SMEX Iowa Satellite Vegetation and Water Index (NDVI and NDWI) Data documentation for information about the atmospheric correction made for NDWI.

Watershed Calibration

VWC was derived from the NDWI, which is a ratio of bands available from the Landsat 5 TM and Landsat 7 ETM+. The ratio is defined as:

Five scenes were used to estimate VWC for the entire watershed. The first step was to determine the trend of VWC versus day of year (DOY) and compare it to the trend of NDWI versus DOY. Vegetation sampling was conducted before and during the SMEX02 to support this endeavor. 31 watershed fields were sampled between 2 and 4 times during June and July 2002 to provide an estimate of VWC per field. Within each field, locations were selected to represent a high, a low, and an average vegetation cover. The corners of each sampling site are the basis for data retrieval for the proceeding analysis.

Average VWC values were calculated for each field for each day of sampling. Some days were combined so that the VWC estimate is the average of five samples at three sampling sites for a total of 15 samples. An initial value of 0.05 kg/m² for VWC was set on June 10, 2002 which is characteristic of the region. From these data points, VWC values were interpolated for the remaining dates for each field. Functions were then developed for both corn and soybean fields.

Average NDWI values were retrieved from the five scenes for each of the WC fields sampled. Field averages were calculated by sampling the four corners of each sampling site for a total of 12 pixels per field. Some pixels were contaminated by nearby roadways or streams and these pixels were eliminated from the analysis. These field averages were then linearly extrapolated between the five days of record to provide an interpolated NDWI product for each field. A relationship was established for NDWI as a function of DOY for the two dominate crop types, corn and soybean.

The DOY functions for both VWC and NDWI were combined to produce new functions where VWC is predicted by NDWI. These functions are as follows:

Corn: VWC = 9.8166 * NDWI + 0.0522
Soybean: VWC= 1.468 * NDWI + 1.3615 * NDWI + 0.3394

In addition, areas determined to be forested were given a VWC of 10 kg/m² and grassland were given a VWC of 0.5 kg/m².

These functions were used to generate VWC maps for the five original TM/ETM+ scenes available. From the five available dates, VWC was determined for the length of the experiment by interpolating on a pixel-by-pixel basis from the five dates of VWC maps. Using a simple piecewise cubic Hermite interpolation for the five days, a time series was established from 6 June through 17 July 2002. This interpolation maintains the pattern for each day of record and interpolates between them for missing days. It was noted that for both NDVI and NDWI, the satellite sensors tended to saturate on or around 8 July 2002 for the very green vegetation. To compensate for this, the values for 17 July 2002 were adjusted in the case of VWC values equal to 5 kg/m³. If a pixel had a value of 5 kg/m² (near saturation) the VWC was increased to 7 kg/m³. This value was used for interpolation.

To verify the accuracy of this method, the pixels relating to the vegetation ground sampling were retrieved for statistical analysis.

Corn: Bias: -0.0097 rmse: 0.5759
Soy: Bias: -0.0153 rmse: 0.1713
Total: Bias: -0.0118 rmse: 0.4695 R²: 0.9302

Regional Calibration

Preliminary regional VWC maps were generated similarly to the watershed, with some exceptions. Three sets of scenes were available for the Iowa region during the experiment, 23 June, 1 July, and 17 July. The two scenes for each day were mosaicked and regional maps of NDWI were generated. From the NDWI maps, VWC maps were generated using the same relationships that were developed in the Walnut Creek (WC) watershed region. From these three scenes, a linear interpolation was developed for the entire study period. To verify the accuracy of the preliminary interpolation method, the WC sampling sites were revisited and statistics were recalculated.

Corn: Bias: -0.2619 rmse: 0.6501
Soy: Bias: -0.0497 rmse: 0.1965
Total: Bias: -0.1883 rmse: 0.5381 R²: 0.9150

The high corn bias can be explained by the inability to interpolate during the later stages of the study period, when there was a precipitation event and rapid growth. This fact coupled with the tendency for NDWI to saturation at approximately 5 kg/m² results in the poor fit of this data.

A primary reason for the preliminary nature of this data is the presence of clouds on 17 July 2002 which potentially influenced VWC readings for that area. Future datasets will provide a better remedy for this issue.