Within the last few decades, forest loss in the Amazon forests has been monitored using satellites such as Landsat (30m resolution) and MODIS Terra and Aqua (250-1000m resolution). Detecting deforestation is relatively easy due to the abrupt changes in the landscape, from vegetation/forest to exposed soil or pasture. This shift causes large changes in the spectral signal (different types of surfaces reflect radiation differently, like its own fingerprint, and is a function of wavelength) measured by the satellite sensors, especially in the near infrared wavelength. The difficulty stands on reliably and systematically assessing the whole Amazon forests (>5.5 million km²) every year in order to guide public policies and action. In this sense, Brazil is a reference for deforestation monitoring through the PRODES program of INPE – the Brazilian National Institute for Space Research (Figure 1). PRODES, allied with another system that produces real-time deforestation alerts (DETER), are the core of the Brazilian efforts on reducing deforestation, with great success during the past decade.
Figure 1 – PRODES deforestation product at TerraBrasilis platform from INPE. Yellow and green represent deforested and forested areas, respectively. Source: TerraBrasilis platform (http://terrabrasilis.dpi.inpe.br/).
However, deforestation is not the only process that removes trees and causes damage to the forests. The harvesting or logging of trees with high economic interest can also cause major forest degradation in the Amazon forests (Figure 2). Although only effectively aimed at specific species and trees, this process kills at least 5 other neighbouring trees when target trees are felled (Johns et al., 1996). Detecting such a subtle process is more challenging than assessing deforestation over a landscape. It requires time series analysis (analysis over a longer time period), very high resolution (VHR, < 1-m) resolution, and at least annual imagery (Dalagnol et al., 2019). Meanwhile, attending to such requirements at larger scale is almost impossible in an environment with high cloud cover. Based on 30m Landsat data, the best estimates of logging impacts that we have consist in 12,000 km²/year of forests being affected in the Brazilian Amazon, producing gross carbon losses of ~0.08 Pg C/year (33% of the Amazon’s annual carbon budget) (Asner et al., 2003; Aragão et al., 2014). However, we do not yet have a precise and generic semi-automatic approach for large-scale degradation mapping and annual monitoring.
Figure 2 – Tree loss represented by very high resolution (VHR, in this case 0.5-m) satellite imagery and airborne LiDAR data. The arrow indicates a tree (with a crown >20 m diameter) that was logged in between the two data acquisition. In the LiDAR data panels, the color represents an increase in height from blue to green/yellow, up to ~40 m. Source: Ricardo Dalagnol.
The use of LiDAR (Light Detection And Ranging) data, loosely translated to “lasers”, can contribute to our understanding on the impacts of forest degradation and improve our ability to detect it using other large-scale products. Small-footprint airborne LiDAR data can precisely detect changes in the vertical forest structure (Figure 3). Although being the best technology around for this kind of observation, the data usually covers only a few square kilometers and data acquisition is still expensive and do not have much availability over tropical forests.
Figure 3 – Visualization of small-footprint Airborne LiDAR data. (A) A LiDAR point cloud observed from above (colors represent height increasing from blue-yellow-red) and a delineated transect. (B) A vertical profile of the delineated transect from panel A. Source: Ricardo Dalagnol.
In our paper published in the Remote Sensing journal (Dalagnol et al., 2019; http://dx.doi.org/10.3390/rs11070817), we specifically investigated the detection of individual tree losses associated with selective logging using VHR satellite imagery and airborne LiDAR data at the Jamari National Forest, Rondônia state, Brazil. This is a national forest with concession for private management through selective logging procedures. We found that logging caused severe changes to the canopy structure which could be detected with >97% accuracy using multi-temporal airborne LiDAR data, and >60% accuracy using multi-temporal VHR satellite data. The gaps in the forests generated by the logging activities were almost all closed after 2 years of regeneration, hindering the detection of this process from the canopy standpoint. It is evident, however, that the damage to the forest structure was already done and the forest would take many decades to regenerate.
Given the great amount of information brought by the LiDAR data, there have been joint initiatives from EMBRAPA/Brazil and U.S. Forest Service to acquire and openly distribute airborne LiDAR data, such as the Sustainable Landscapes Project (https://www.paisagenslidar.cnptia.embrapa.br/webgis/), and the Amazon Biomass Estimate (EBA) project from Dr. Jean Ometto and his team from INPE (http://www.ccst.inpe.br/projetos/eba-estimativa-de-biomassa-na-amazonia/). The EBA data is a recent initiative and consist of 1000 flight lines acquired during 2017-2018, where at least 624 of those are non-overlapping and distributed across the whole Brazilian Amazon (Figure 4), and the remaining may be repeated measurements over areas of interest for forest degradation studies. Recent acquisition of large-footprint orbital LiDAR data from the Global Ecosystem Dynamics Investigation (GEDI) onboard the International Space Station (ISS) also bring priceless data across tropical forests, which will contribute to better understanding the forest structure and changes to it. I expect great advances in the understanding of tropical forests in the upcoming years derived from these datasets.
Figure 4 – Distribution of airborne LiDAR transects from the EBA/INPE project (n = 624). Source: Ricardo Dalagnol.
Integrating all the available large-scale satellite data to develop a detection method has never been so easy. Advances in image processing brought by cloud computing platforms, such as the Google Earth Engine, readily distribute the whole portfolio of several satellite imageries. For example, all series from Landsat satellite (1985-today) with 30-m spatial resolution, and Sentinel-2 satellite (2014-today) with 10-m spatial resolution, can be used to generate global maps. The imagery can be manipulated using several programming languages (e.g. R, Python) through the Google Colab platform.
Moreover, scientists can use advanced artificial intelligence methods for data analyses such as Deep Learning models available in R and Python (e.g. through Keras and TensorFlow backends). One example of this kind of model is the U-Net (Figure 5), which is a deep learning method for pattern recognition and image segmentation, and have been used, for example, for tree species detection (>95% precision) using very high resolution satellite imagery (Wagner et al., 2019). Such model could be taught to detect the patterns caused by logging, such as opening of trails and roads, treefalls, slightly opening the canopy, and be applied for large-scale mapping using the 10 to 30-m spatial resolution data (Landsat and Sentinel-2).
Figure 5 – Deep learning U-net model architecture. Source: adapted from Wagner et al. (2019).
Therefore, integrating the knowledge from coupling LiDAR data with large-scale satellite data (Landsat / Sentinel-2) through state-of-the-art Deep Learning models seem to be a way forward on large-scale assessments of Amazon forests degradation.
About the Author: Dr. Ricardo Dalagnol is a Remote Sensing scientist specialized in vegetation and forest structure. His latest work focuses on tree mortality and forest disturbance in the Amazon forest. He is currently working as post-doc researcher at the National Institute for Space Research (INPE) in Brazil. For more information you can contact him by email (firstname.lastname@example.org) or follow him at Twitter (@RicardoDalagnol).
Asner, G.P.; Knapp, D.E.; Broadbent, E.N.; Oliveira, P.J.C.; Keller, M.; Silva, J.N. Selective Logging in the Brazilian Amazon. Science 2005, 310, 480–482.
Aragão, L.E.O.C.; Poulter, B.; Barlow, J.B.; Anderson, L.O.; Malhi, Y.; Saatchi, S.; Phillips, O.L.; Gloor, E. Environmental change and the carbon balance of Amazonian forests. Biol. Rev. 2014, 89, 913–931.
Dalagnol, R.; Phillips, O.L.; Gloor, E.; Galvã, L.S.; Wagner, F.H.; Locks, C.J.; Aragão, L.E.O.C. Quantifying Canopy Tree Loss and Gap Recovery in Tropical Forests under Low-Intensity Logging Using VHR Satellite Imagery and Airborne LiDAR. Remote Sens. 2019, 11, 817.
Johns, J. S.; Barreto, P.; Uhl, C. Logging damage during planned and unplanned logging operations in the eastern Amazon. Forest Ecology and Management, 1996, v. 89, n. 1–3, p. 59–77.
Wagner, F. H., et al. Using the U-net convolutional network to map forest types and disturbance in the Atlantic rainforest with very high resolution images. Remote Sens in Ecology and Cons., 2019, 1–16.