High-resolution residual dry matter (RDM) map for a California oak savanna/annual grassland derived from drone multispectral remote sensing imagery and in-situ grass biomass data

This dataset is a high-resolution (60 cm) grass biomass map derived from drone/UAS (unmanned aerial system)-based multispectral remote sensing, calibrated to in situ field data. It was developed for research on prescribed fire behavior responses to vegetation conditions, and vegetation community regrowth-responses post-fire. It addressed a key need for nondestructive grassland biomass measurements, in a use case where directly measuring grass biomass by destructive harvest would have disturbed the intact fuel beds needed for burning in the prescribed fire experiment.

Methods
Grass biomass map data collection and processing consisted of three steps: (1) drone flights for multispectral image acquisition, (2) clipping and measuring grass biomass in 1-foot square (0.30 m x 0.30 m), spatially referenced plots, and (3) development and application of predictive relationships between in-situ grass biomass and spectral indices from the drone imagery.  The data object published here is a grass biomass map from 7 October 2020 (2020-10-07) calculated using the predictive relationship developed in step3. 

Multispectral drone imagery was collected using a Draganfly Commander unmanned aerial system (UAS) (Draganfly Innovations, Saskatoon, SK CA) flying a Micasense Rededge multispectral camera (AgEagle, Wichita, KS USA). A mapping flight for developing grass biomass calibrations was conducted over an experimental grassland site at UCSB Sedgwick Reserve, Santa Barbara, CA (34.702, -120.036) on September 16, 2020 (2020-09-16). The Rededge Camera was flown at 90 m above ground level in a pre-programmed grid flight with 80% front and side overlap between images. Multispectral imagery was calibrated to reflectance with before- and after-flight reference panel images and processed into geotiff (.TIF) image mosaics in Pix4D software (Pix4D, Prilly, CH). Image mosaics consisted of five wavelength bands: blue (475 ± 20 nm), green (560 ± 20 nm), red (668 ± 10 nm), red-edge (717 ± 10 nm nm) and near-infrared (NIR) (840 nm ± 40). Raw drone imagery produced had a native spatial resolution of 0.06 m (6 cm).  Imagery was georeferenced to less than 1 m absolute accuracy with 12 in-scene ground control points (GCPs) surveyed with a Trimble PG200 RTK antenna (Westminster, CO USA). 

Next, in-situ grass biomass data (or residual dry matter (RDM) at this dry-season time of year) was collected at 25 locations across the experimental grassland site that represented a range of grass cover conditions. Grass sampling occurred on 25 September 2020, the week after drone image acquisition. Sampling locations were referenced by tape measure to 4-foot T-posts visible in the drone imagery. Grass was clipped to mineral soil and massed within standard-sized, 1-foot square sampling frames used for rangeland monitoring in California annual grasslands (Bartolome et al., 2006). As grass sampling was done after the calibration image flight, the drone imagery thus represented grass biomass conditions at clip-plot sampling locations before clipping. This enabled direct comparison of spatially aligned drone reflectance data and grass biomass. No significant meteorological events such as rain or wind, or additional major physical disturbances occurred onsite between drone mapping and grass biomass sampling.

After sampling grass biomass in the field, step-wise multiple linear regression was used to develop predictive relationships among spectral indices calculated from the Sep 16 drone imagery and field-measured grass biomass (RDM). Regression analyses were carried out in R software (R Core Team 2023, Vienna, AT).  Area-based grass biomass units modeled were pound per acre (lb acre-1) to make the model relevant for Santa Barbara County rangeland managers who track grassland biomass in these units.  The strongest relationship found among field-measured grass biomass and spectral indices was the following (Eq. 1)

Eq. 1:  RDM (lb acre-1) = -52501 - 46600*OSAVI + 72132*TNDVI – 114995*blue + 91390*green

where spectral data and indices calculated from drone imagery included:

OSAVI (Optimized Soil-Adjusted Vegetation Index) =  (NIR-Red)/(NIR+Red+0.16) (Fern et al., 2018; Rondeaux et al., 1996)

TNDVI (Transformed Normalized Difference Vegetation Index) =  √ [(NIR-Red)/(NIR+red) + 0.5] (Gholami Baghi & Oldeland, 2019)

Blue = blue reflectance band

Green = green reflectance band

The relationship in Eq. 1 represented a significant predictive relationship for dry season grass biomass (RDM) from multispectral drone image spectral data and indices (r2 = 0.501, F = 7.23 (4,21), p < 0.001; root mean square error (RMSE): 556 lb acre-1 on 21 D.F.).

The grass biomass map published herein was produced from drone imagery flown on 2020-10-07.  After image collection and processing according to step (1) above, the October 7 imagery was histogram-normalized to the 2020-09-16 imagery to account for differences in sun angle and illumination between dates.  From Step 3, Eq. 1 was then applied to the 2020-10-07 imagery to produce a dry grass biomass/RDM biomass map.  Unfortunately, there was not time to ground-truth the biomass map with independent grass clip plots before the prescribed fire experiment that this map supported. 

References

Bartolome, J., Frost, W., & McDougald, N. (2006). Guidelines for residual dry matter on coastal and foothill rangelands in California. Rangeland Management Series, 092, 1–6.
Fern, R. R., Foxley, E. A., Bruno, A., & Morrison, M. L. (2018). Suitability of NDVI and OSAVI as estimators of green biomass and coverage in a semi-arid rangeland. Ecological Indicators, 94, 16–21. https://doi.org/10.1016/j.ecolind.2018.06.029
Gholami Baghi, N., & Oldeland, J. (2019). Do soil-adjusted or standard vegetation indices better predict above ground biomass of semi-arid, saline rangelands in North-East Iran? International Journal of Remote Sensing, 40(22), 8223–8235. https://doi.org/10.1080/01431161.2019.1606958
Rondeaux, G., Steven, M., & Baret, F. (1996). Optimization of Soil-Adjusted Vegetation Indices. Remote Sensing of Environment, 55, 95–107.