Coupling remote sensing and eDNA to monitor environmental impact: A pilot to quantify the environmental benefits of sustainable agriculture in the Brazilian Amazon

Abstract

Monitoring is essential to ensure that environmental goals are being achieved, including those of sustainable agriculture. Growing interest in environmental monitoring provides an opportunity to improve monitoring practices. Approaches that directly monitor land cover change and biodiversity annually by coupling the wall-to-wall coverage from remote sensing and the site-specific community composition from environmental DNA (eDNA) can provide timely, relevant results for parties interested in the success of sustainable agricultural practices. To ensure that the measured impacts are due to the environmental projects and not exogenous factors, sites where projects have been implemented should be benchmarked against counterfactuals (no project) and control (natural habitat) sites. Results can then be used to calculate diverse sets of indicators customized to monitor different projects. Here, we report on our experience developing and applying one such approach to assess the impact of shaded cocoa projects implemented by the Instituto de Manejo e Certificação Florestal e Agrícola (IMAFLORA) near São Félix do Xingu, in Pará, Brazil. We used the Continuous Degradation Detection (CODED) and LandTrendr algorithms to create a remote sensing-based assessment of forest disturbance and regeneration, estimate carbon sequestration, and changes in essential habitats. We coupled these remote sensing methods with eDNA analyses using arthropod-targeted primers by collecting soil samples from intervention and counterfactual pasture field sites and a control secondary forest. We used a custom set of indicators from the pilot application of a coupled monitoring framework called TerraBio. Our results suggest that, due to IMAFLORA's shaded cocoa projects, over 400 acres were restored in the intervention area and the community composition of arthropods in shaded cocoa is closer to second-growth forests than that of pastures. In reviewing the coupled approach, we found multiple aspects worked well, and we conclude by presenting multiple lessons learned.

Fig 1. An overview of the proposed coupled approach to environmental monitoring.

In the remote sensing component, we identify and pre-process key satellite imagery. Next, we input data into change algorithms or other models. Then, we assess the accuracy of the produced maps. In the eDNA component, we design the sampling approach by identifying locations to sample. Next, we visit sites to collect eDNA data. Then, the soil samples go through the eDNA extraction and biodiversity analysis processes. Finally, we calculate indicators from the map outputs and biodiversity results.

Fig 2

Shade grown cocoa agroforestry in the Xingu River basin begins with cultivating shade trees in cocoa agroforestry systems, alongside annual crops and young native trees (left). Over time, the cultivated shade trees are thinned and the native shade trees become dominant (right). Created using components from the University of Maryland Center for Environmental Science Integration and Application Network media library; components copyright Tracey Saxby, Jane Thomas, and Dieter Tracey; under Attribution-ShareAlike 4.0 International.

Fig 3. Study area in the state of Pará, Brazil.

Yellow dots represent the farms that have partnered with IMAFLORA.

Fig 4

(A) Disturbance classification 2010–2015. Inset maps (B: B1, B2, B3) are shown with black outline and inset maps (C: C1, C2) are shown with blue outline on the main map. (B1) Inset map of disturbance classification 2010–2015 over a particular area to outline classification vs. RGB images. (B2) Inset RGB image pre-study period from Landsat 5 (July 30, 2009). (B3) Inset RGB image post-study period from Sentinel-2 (June 26, 2016). (C1, C2) Example of disturbances within one of the properties (degradation in yellow, deforestation in red, stable forest in dark green, and non-forest in gray). overlayed on the RGB Landsat 5 image (July 30, 2009).

Fig 5

(A) regeneration map output 2010–2015 overlayed on Sentinel-2 imagery (ESA). Inset maps (B: B1, B2, B3) are shown with black outline and inset maps (C: C1) are shown with blue outline on the main map. (B1) Inset map of regeneration output 2010–2015 over a particular area to outline regeneration areas vs. RGB images. (B2) Inset RGB median image pre-study period from Landsat 5 (July-August 2011). (B3) Inset RGB image post-study period from Planet NICFI mosaic (July—November 2016). (C1) Example of regeneration areas (in green) within two of the properties.

Fig 6. Species richness by field type.

Each dot represents a sample point in shaded cocoa fields (red) or pasture fields (blue).

Fig 7

Shannon diversity (left) and Simpson’s (right) by field type.

Fig 8. Effective species richness (L) and Inverse Simpson Diversity (R) by field type.

Each dot represents a sample point in shaded cocoa fields (red) or pasture fields (blue).

Fig 9. Beta diversity (Aitchison distance) between each site.

Blue corresponds to lower dissimilarities between sites and yellow corresponds to higher dissimilarities between sites.

Fig 10. Aitchison distances within treatments and between treatments.

Fig 11. Arthropod PCA for arthropods found in shaded cocoa, forest, and pasture sites.

Fig 12. Plot in the CEO disturbances validation project.

This pixel was misclassified as Degradation. NDFI time series show a seasonal pattern with lower NDFI values around August of each year. It is important to note that the MapBiomas classification maps these areas as “Savannic Forest Formations”. Basemap imagery is from Sentinel-2 (European Space Agency).

Fig 13. Example plot of the CEO change validation project showing a fragmented/degraded forested patch and its variations on the NDFI time series.

Basemap imagery is from Sentinel-2 (European Space Agency).

Fig 14. Example plot of the CEO change validation project showing a plot that was classified as regrowth by the algorithm, but the interpreter classified it as a single degradation event.

We note some regreening between the dates and the variation in NDFI values. Although not entirely clear, the interpreter assumed selective logging followed by a fire event in 2011. Another fire event seems to have happened in 2012. It is not clear that by the end of 2014 the area was already-established pastureland.