Reports

Contents

Title: Tonzi Tower - NDVI (SRS, Planet)
Date:2017-09-11 - 2021-05-04
Data File: TZ_NDVI.csv
Refers to:TZ,2046503368,1066703767,943103015

Various sensors measuring NDVI have been installed across our sites over the last several years. At some sites, the SRS sensors from Decagon/METER have been drifting low, so I wanted to compare them with other NDVI measurements.

  • SRS sensor (Decagon/METER): A pair of photodiodes with cosine correcting Teflon diffuser and hemispherical view. Red band is measured at 650nm and NIR band is measured at 810nm.
  • Planet Labs data: L3H (CESTEM) alpha data product. Pixel size is 3m, and there are daily gap-filled images from 2018-2020. Gap fills are done with Planet imagery from the day before and after, combined with daily MODIS data. From the full image, Joe cut out a 730mx730m tile for each site, centered on the tower. Product includes NDVI, GCC, ECI, and the following reflectances: red, NIR, green, blue. For more information, see Box/Biometlab/Remote_Sensing/PlanetLabs/L3H_distribution_lodi_islands_readme.pdf.
  • Broadband NDVI: NDVI calculated using reflected SW and PAR data from the tower, called "broadband" because it uses the broader bands of shortwave and PAR radiation rather than the narrow bands of the SRS and Apogee sensors. See Huemmrich et al., 1999 in Journal of Geophysical Research: Atmospheres and Tittebrand et al., 2009 in Theoretical and Applied Climatology.

I used Joe's datafetch tool to calculate daily average mid-day values of NDVI from the various sensors. Mid-day values included data from 11:00 to 13:00, 5 values a day. I despiked the data in Excel.

  • NDVI = (NIR-Red)/(NIR+Red)
  • Broadband NDVI = (rho_NIR-rho_PAR)/(rho_NIR+rho_PAR)
    • rho_NIR= (SWout-PARout)/(SWin-PARin)
    • rho_PAR=(PARout/PARin)
    • Units of both SW and PAR should be in W/m2. Use the conversion 4.6 umol/m2/s = 1 W/m2 to convert the usual PAR units (umol/m2/s) to W/m2.

 

Site SRS sensor (incoming/outgoing) PAR sensor (incoming/outgoing)
 4-way radiometer
 Conclusion
Tonzi Tower

 

 

 sn 010164/sn 030382

sn 020476
  • 650in bad starting 2020.
  • 810out (hemispherical) most likely bad. 
  • Other bands maybe ok.
  • Incoming SRS data and outgoing grass data should be okay through 2019.

NDVI

Figure 1. Tonzi Tower NDVI. Both SRS and broadband NDVI seasonal peaks in 2020 are lower than in 2018 and in 2019, while PL NDVI seasonal peak is about the same. Maybe the difference between broadband and PL peaks can be explained by their different sampling areas?

 

Individual Red/NIR bands

 

Figure 2. Individual incoming bands of SRS sensor. 650in bad starting 2020 but 810in looks ok.

 

Figure 3. Individual outgoing bands (hemispherical sensors) of SRS sensor. 650out seasonal peak is lower in 2019/2020 than in 2018. 810out seasonal peak is lower in 2020 than in 2018/2019. Not sure yet if this is real observations or if it's sensor decay.

 

Figure 4. Individual outgoing bands (field-stop sensors pointed at grass) of SRS sensor. 650out and 810out bands both decrease year after year, but hard to tell if it's real signal or if it's sensor decay.

 

Red reflectance

Figure 5. Red reflectances from SRS and PL data. Grass and PL reflectances reach their annual minimum around the end of April, while the hemispherical SRS reflectance reaches its annual minimum about 4-6 weeks later.

 

Regression Data

Residuals

Figure 6. Linear regression of red reflectance between SRS and PL data. There are multiple clusters in the regression between hemispherical SRS and PL reflectances, which indicates either 650in or 650out is bad. We know from the timeseries figure that 650in was bad starting early 2020, but it's possible 650out is also bad.

 

NIR reflectance

Figure 7. NIR reflectances from SRS and PL data. Hemispherical SRS seasonal peaks decrease year over year compared to PL seasonal peaks, which indicates either 810in or 810out is bad. Since the 810in seasonal peaks have been consistent over the past few years (Fig 2), 810out is probably bad starting in 2019, unless we think the difference can be explained by sampled area.

 

Regression Data

Residuals

Figure 8. Linear regression of NIR reflectance between SRS and PL data. R2 < 40%; poor relationships all around.

 

NDVI regressions with SRS (hemispherical)

Regression Data

Residuals

Figure 9. 2017-2020 data: Linear regression of NDVI between SRS (hemispherical), broadband, and PL data. Messy.

 

Regression Data

Residuals

Figure 10. 2017-2019 data only: Linear regression of NDVI between SRS (hemispherical), broadband, and PL data. This regression looks cleaner without 2020 data (R2 increased from ~0.23 to ~0.5), but it looks like there's still hysteresis in the relationship.

 

Regression Data

Residuals

Figure 13. 2017-2018 data only: Linear regression of NDVI between SRS (hemispherical), broadband, and PL data. R2 has increased from ~0.5 to ~0.7, but still have obvious hysteresis.

 

NDVI regressions with SRS (field-stop)

Regression Data

Residuals

Figure 11. 2017-2020 data: Linear regression of NDVI between SRS (field-stop), broadband, and PL data. Also messy.

 

Regression Data

Residuals

Figure 12. 2017-2019 data only: Linear regression of NDVI between SRS (field-stop), broadband, and PL data. Looks much cleaner without 2020 data (R2 increased from ~0.7 to ~0.92). Incoming SRS data and outgoing grass data should be okay through 2019.