Using multi-spectral phenological trajectories to detect an invasive grass in the Sonoran Desert

Background/Question/Methods

Since many plants in arid lands have unique phenological attributes, spectral vegetation indices can be used to identify differences among individual species. Moreover, biological invasions may induce phenological changes that result in novel opportunities for remote-sensing based detection. Although most techniques seek distinct signals in Normalized Difference Vegetation Index (NDVI) time series, these approaches do not leverage the full potential of multispectral data. This is especially true in arid lands where photosynthetically active vegetation comprises a small fraction of the landscape for most of the year. We tested this hypothesis using the invasive C₄ bunchgrass, Cenchrus ciliaris, which threatens the Sonoran Desert by introducing a grass-fire cycle. We compared the power of Landsat TM bands and normalized TM band differences to discriminate abundance of buffelgrass in Sonoran Desert uplands in southern Arizona. We converted 104 cloud-free TM scenes collected between 2000 and 2010 to ground-level reflectance and measured the performance of each band (1-5,7) and band ratio ((b2-b1)/(b2+b1) for all 15 possible combinations of two TM bands) using buffelgrass counts at 35 Landsat TM pixel-sized plots (30x30m) and simple linear regression models. We tested each band and ratio on individual scenes and pairs of sequential scenes by comparing adjusted R² values of the 21 single season and 21 multi-date regression models.

Results/Conclusions

The most discriminating band for individual scenes and sequenced scenes was band 2 (green), followed by band 7 (shortwave infrared region 2, or SWIR2) and band 4 (near infrared, or NIR). Green explained more variance than all other multi-band indices, including NDVI. NDVI was consistently outperformed by six band combinations: 7/5, 4/2, 3/2, 7/4, 7/3, and 5/3, four of which included a SWIR band and 3 of which included red. Although the NIR-based indices NDVI (4/3) and 4/2 performed better following the monsoons in 2002, 2003, and 2006, SWIR-based indices and the green band were consistently better during the arid foresummer (April-June) and post- monsoon senescence (October-December). The most discriminating models were based on scenes acquired during the arid foresummer when NDVI is typically low. We surmise that functional change induced by buffelgrass invasion (e.g., increase in non-photosynthetic vegetation replacing bare ground) is responsible for the stronger relationship with SWIR and plant pigments are responsible for the power of visible bands. This study underscores the need for multispectral phenological assessments of invasion and highlights the fact that NDVI alone may not discriminate  functional and structural changes.