1. Bamboo-dominated forests (BDF) extend over large areas in the drought-prone Southwestern Amazon, yet little is known about the dynamics of these ecosystems. Here, we investigate the hypothesis that bamboo modulates large-scale ecosystem dynamics through competition with coexisting trees for water.

2. We examined spatio-temporal patterns of remotely sensed metrics (Enhanced Vegetation Index [EVI], Normalized Difference Moisture Index [NDMI]) in >300 Landsat images as proxies for canopy leaf phenology and water content at two time scales: (1) a complete bamboo life cycle (~28 years), and (2) the seasonal cycle; and at two spatial scales: (a) comparing adjacent areas of BDF vs. Terra-firme forests (TFF) to investigate regional dynamics, and (b) comparing the vegetation classes of bamboo, trees in BDF, and trees in TFF to investigate the effects of bamboo on coexisting trees.

3. At the regional scale, BDF showed higher EVI (leaf area density) and lower NDMI (water content) than nearby TFF but these differences disappeared as bamboo died, suggesting a strong influence of bamboo life-stage in the functioning of these forests. BDF seasonal cycle showed a bimodal EVI pattern as trees and bamboos had asynchronized leaf production peaks.

4. At the scale of vegetation classes, trees in BDF showed lower NDMI (i.e., water content) than trees in TFF except after bamboo mortality, indicating a release from competition with bamboo for water. Canopy water content of trees in BDF was also reduced during bamboo dry-season greening (increased EVI ~ leaf production) due to increased water demands. Nevertheless, long-term and seasonal phenology of trees in BDF did not differ from that of trees in TFF suggesting a potential selection for drought-tolerant trees in BDF.

5. Synthesis. Bamboo-dominated forests have received less attention than other Amazonian forests and their functional dynamics are commonly ignored or misinterpreted. Using remote sensing to characterize forest phenology and water content, we show the distinctive seasonal and long-term dynamics of BDF and coexisting trees and the importance of bamboo competition for water in shaping this ecosystem. Our results suggest a potential selection for drought-tolerant trees in BDF since they maintain the same EVI as trees in bamboo-free forests but with lower water content. A better characterization of BDF and their cyclical dynamics is crucial for accurately interpreting Amazonian forests’ responses to extreme climatic events such as high temperatures and droughts.

We bulk downloaded 964 multi-spectral Landsat acquired images from the USGS Earth Resources Observation and Science (EROS) Center Science Processing Architecture (ESPA) (https://espa.cr.usgs.gov/). The 964 images had been acquired by Thematic Mapper (TM), Enhanced Thematic Mapper (ETM), and Operational Land Imager (OLI) over the study area between July 1987 and August 2018. In order to reduce the effect of the atmosphere on electromagnetic radiation when it traverses earth’s atmosphere, we obtained atmospherically corrected Landsat Collection 1 surface reflectance products together with quality layers, cloud masks and reflectance-derived spectral vegetation and water-content indices. Surface reflectance and index data layers were masked using the pixel QA band, which includes probability estimates for clouds and cloud shadows. We only kept high-probability cloud-free terrain pixels (clear terrain value 66 for OLI, value 322 for the other sensors). In the following steps, only images from 1989 onwards (n = 901) were included. Surface reflectance and index data layers were masked using the pixel QA band, which includes probability estimates for clouds and cloud shadows. We only kept high-probability cloud-free terrain pixels (clear terrain value 66 for OLI, value 322 for the other sensors). In the following steps, only images from 1989 onwards (n = 901) were included.

In order to study the seasonal and inter-annual patterns of BDF and TFF, we randomly selected 10 sets of 5,000 pixels each within each of the two delineated study polygons (~4.5% of pixels in each polygon per set) using a cloud-free Landsat image (LC800406820130830). For each of the randomly selected pixels, we extracted EVI, NDMI, and NDVI values from the data cubes and calculated means and standard deviations for each set, polygon (TFF and BDF) and image.

Although the previous cloud-masking procedure eliminated the majority of clouds and cloud-shadows, some image dates still showed extreme values that corresponded to unmasked clouds and shadows within the study area polygons. To correct for this, we manually checked the images with NDVI means between 0 and 0.75 or EVI means between 0 and 0.45 to corroborate the presence of clouds, cloud-shadows or haze. Then, we eliminated all of the images with set means of NDVI >1 or NDVI <0.7 or EVI >1 or EVI <0.4 as they were found to contain large areas with clouds. We also removed two images from 2012 that showed outlier values in the time series due to partial cloud cover. After filtering, 350 of the 901 original images remained for use in the regional-scale analysis. 

To disentangle the contribution of bamboo, trees in BDF, and trees in TFF to regional-level reflectance patterns, and to investigate the response of trees to bamboo co-existence, we analysed a subset of manually designated pixels within each of the two study polygons. First, we extracted the spectral signatures (all bands) of the three vegetation classes from pixels that were predominantly bamboo, trees in BDF and trees in TFF from a cloud-free Landsat image (LC800406820130830) based on the reference points identified in the GeoEye image. On the basis of their spectral signatures, we digitized a set of 119 pixels representing areas with bamboo, 286 pixels of trees within BDF, and 279 pixels of trees in nearby TFF with no observable bamboo presence.

For each of the 684 selected focal pixels, we extracted EVI, NDMI, and NDVI values from the same collection of 350 images used for the regional analysis. We then calculated the mean pixel values per class (bamboo, trees in BDF, and trees in TFF) and image for each vegetation index. We further eliminated images when index means were calculated from <10 pixels (~0.9 ha). This threshold retained most of the data while removing data points that could be highly influenced by just a few pixels with remaining clouds, shadows, or other artefacts.

We extracted image acquisition date for each image and then used a moving window of 360 days (~1 year) for the regional analysis and 180 days (~half year) for the individual-based analysis. The wider moving window offers a smoother (deseasonalised) trend suitable for long-term regional analysis while the narrower window captures seasonal variation. For the individual analysis, filtering removed all but two of the images from 1992; as such, we removed all dates between 1991-10-15 and 1993-01-15 from this analysis. For the values within the moving window we calculated mean and 95^th quantile confidence intervals around the bootstrapped median (time-series bootstrap with stationary block resampling where block length has a geometric distribution with a mean of half of the count of the number of observations included within that window, 1,500 iterations, R package “boot”) (Buckland, Davison, & Hinkley, 1998; Canty & Ripley, 2017; Politis & Romano, 1994). At the individual scale, we obtained a separate series for each vegetation class (bamboo, trees in BDF and trees in TFF). At the regional-scale analysis, we obtained a separate series for each of the 10 sets per forest type and the 95% confidence intervals of their means.