Ecosystem Insights from Space

Satellite imagery analysis has emerged as a powerful tool for understanding our natural environment. This technology enables ecosystems to be analyzed at large spatial scales at a relatively low cost, to address critical issues such as vegetation health, landcover change, and habitat connectivity. This technology is another powerful and exciting tool that AJM has available for informing responsible environmental management.

Seeing the Invisible

The world is a lot more colorful than our human eyes can recognize. For example, in Fall we see the leaves change from green to shades of yellow, orange, and brown (i.e. less green light and more red light is being reflected from leaves), but we don’t see that these leaves are now reflecting much less near-infrared light than they were throughout the summer. Modern satellites, however, capture data from a far wider range of wavelengths than our eyes can process. Being able to detect these wavelengths beyond the visible spectrum is incredibly valuable, as it provides deeper and more precise insights into environmental processes that would otherwise remain hidden (Figure 1). The possible applications of this data are vast, but here are some of the techniques that AJM has recently used.

Figure 1: By displaying reflectance from wavelengths beyond our visual spectrum as a color that we can perceive, we visualize these signals. Here, we can see the near-infrared reflectance, where dark red represents low reflectance, and white represents high reflectance.

Vegetation Indices

Vegetation indices are one of the simplest yet powerful applications of satellite imagery. These indices combine two or more wavelengths to quantify the biophysical condition of vegetation. The most prominent vegetation index is the Normalized Difference Vegetation Index (NDVI), which is commonly used to measure vegetation health. This index leverages the fact that healthy vegetation reflects a lot of near-infrared light and not much red light, where unhealthy vegetation reflects a lot of red light and not much near-infrared light. Using a simple normalizing equation, all pixels in a satellite image can be scored from -1 to 1, where 1 is very healthy vegetation and anything below zero is devoid of living vegetation. We used this technique to measure wetland vegetation health in a boreal forest (Figure 2), but this analysis can also be used to monitor crop growth, detect vegetation diseases, map drought vulnerability, and estimate wildfire fuel loads, among other uses (Huang et al. 2021). Indices are not limited to vegetation, however, as similar techniques are available to measure environmental characteristics such as soil moisture, snow cover, and water quality. 

Figure 2: NDVI analysis, where light green represents low scores (less health vegetation, or no vegetation), and dark green represents high scores (healthier vegetation).

Machine Learning Analysis

Machine learning is one of the most exciting new frontiers in satellite imagery analysis. Machine learning techniques all utilize pattern recognizing algorithms that would be prohibitively time-consuming for humans to complete without computer assistance. These techniques are able to consider a wide range of spatial data and make predictions based on complex relationships existing within the data. We have used two different machine learning techniques to analyze satellite imagery.

We used a ‘random forests’ classification to predict where stressed and dead vegetation may be occurring. Random forests uses a group of decision trees (the “forest”) to make predictions. Each decision tree in the forest is created using a random subset of the training data (a process called ‘bagging’), so each tree in the forest is different. This introduces variability between trees and reduces the risk of overfitting. In classification, each tree in the forest will individually “vote” for which class it thinks the data belongs to, and a classification will be based on the class for which the majority of trees vote (Breiman 2001). We used this technique to classify landcover in a boreal forest and predict where unhealthy vegetation may be occurring (Figure 3, left). Our analyses combined satellite imagery (red, blue, green, and near-infrared wavelengths) with spatial data describing wetland extent and an NDVI layer. Using this technique, we were able to predict the presence of unhealthy vegetation with high accuracy.

Another technique we used alongside random forests to detect stressed and dead vegetation was gradient boosting. Similar to random forests, gradient boosting employs an ensemble of decision trees, but it builds these in sequence. Each subsequent tree is trained on the residual errors of the previous trees, with its contribution to the final prediction being progressively smaller in a process known as ‘boosting’ (Elith et al. 2008). We applied gradient boosting to the same input data as the random forests analysis, but instead conducted a regression analysis to predict the likelihood of stressed or unhealthy vegetation occurring across the landscape (Figure 3, right). This technique can enable us to identify and prioritize areas of potential concern to inform targeted management actions.

Figure 3: A random forests classification (left) showing water (blue), low vegetation (light green), forest (dark green), and unhealthy vegetation (yellow), and a gradient boosting regression analysis (right) showing the probability of unhealthy vegetation being present (darker red represents a higher probability).

The field of satellite imagery analysis is rapidly evolving, driving advances in sensor technology, analytical techniques, and computing power. As this evolution continues, we can expect even greater potential for addressing complex ecological questions at large scales. At AJM, we are excited to be part of this journey, continuously adapting and applying the latest satellite imagery analysis methods to better understand and manage the landscapes we love.

Written by: Simon Coats, BSc., MGIS - AJM Environmental Scientist

References

Breiman L. 2001. Random forests. Mach. Learn. 45:5–32.

Elith J, Leathwick JR, Hastie T. 2008. A working guide to boosted regression trees. J. Anim. Ecol. 77:802–813.

Huang S, Tang L, Hupy JP, Wang Y, Shao G. 2021. A commentary review on the use of normalized difference vegetation index (NDVI) in the era of popular remote sensing. J. For. Res. 32:1–6.