The remote sensing case for small-scale tropical agroforestry
Challenges in Optical Remote Sensing and how to address them
In my day-to-day as an image analyst I am often objected to the idea that everything can be seen from space. It is an incredible source of truth with enormous potential — independent from any agendas, just a quiet pass every few days to reveal what happens on planet Earth.
And yes, it is true. We all have been getting used to the earth-from-above type images that are easy to the eye. But, there are limits to what can be seen and challenges in how to work around it.
Let’s explore the three biggest challenges for Optical Remote Sensing in tropical (very) small-scale agroforestry and how to address them. Well, frankly its already in the name: Tropical, small and agro-forestry …
3 main challenges:
1. Cloud cover
Problem:
Operating in the tropics means mostly one thing to Remote Sensing: cloud cover. And while we often associate rain forests with large cumulonimbus type clouds that bring heavy rains, we experience as often cotton-like cumulus clouds that you might recognize from a warm and pleasant summer day. So that in reality we have to deal with persistent cloud cover in wet season, and the more-than-likely cloud covering just the spot that you are interested in within dry season — despite good weather.
Limitation:
This is limiting for the use of optical Remote Sensing that cannot penetrate clouds. Not being able to take pictures of the earth means we cannot monitor the plot at a given moment in time. But it also means that creating composites is getting harder to do as they become more unreliable and uglier to look at.
Solution:
In short: We cannot make conclusions on what hasn’t been observed.
But,..
One can become quite good at “guessing” what’s below the clouds with filling algorithms / modelling, an exercise often necessary for Near-Real-Time applications or products that require sums (e.g. Total Biomass or Yield products). We differentiate between 2 types of filling: spatial (neighborhood-based) and temporal (pixel-based) /or a mix of both.
One can make better guesses using real observations from other sensors. Other optical sensors might have better temporal resolution, which means they come over more often and therefore increase the chance of observation. This solution comes in trade-off with lower spatial resolution and is addressed with Data Fusion algorithms.
One can also simply accept the data absence by choosing a higher level of aggregation, e.g. monthly or yearly maps. Ideally, these maps are accompanied by a quality map to explain the different amount of input images per location, so that the user can be aware that in certain areas their “yearly” value is e.g. entirely based off the one observation in January.
Radar sensors can penetrate clouds and are the preferred solution for scenarios of very dense and persistent cloud cover. Their trade-off up until now was a decreased pixel resolution (e.g. Sentinel 1 at 20m instead of Sentinel2 at 10m). But with the latest satellite launches such as NISAR (5m), they are a very good alternative for applications such as deforestation monitoring. However, for certain applications that actually need the spectral information from the optical sensor (e.g. Land cover classification), Radar is no good alternative.
2. Plot size vs pixel size
Problem:
Small-scale agriculture refers to “small” field sizes. There is no common definition on how small or big a “small-scale” agricultural plot really is. From what I gather from my searches on the internet it seems context-based and depends on the commodity and country. The examples I have in mind are VERY small-scale at 0.2–2 ha based on my work in planting trees in coffee and cocoa plantations.
The small size in itself is not really the problem. What is though, is its relation to pixel size.
Limitation:
Popular, because free, optical sensors operate at 10m (Sentinel-2) or 30m (Landsat) resolution. And also freely available Radar sensors operate at similar resolutions (Sentinel-1, PALSAR 20m). Individual trees are not visible at this scale.
For best results it is preferable to have object sizes of min. 8 -11 times the pixel size (Lechner et al. 2009, Vajsová et al. 2020). This means that a field should at least be 80x80 m (0.64 ha) to be mapped with Sentinel-2, or 240x240 m (5.76 ha) to be mapped with Landsat. If this is not the case we are mostly dealing with so-called “mixed pixels” that will largely influence the result. A mixed pixel contains multiple smaller objects and that will present a sum of characteristics by all objects contained within it.
Other than the overall size, the shape can also be limiting for the same reason. Elongated features have a higher proportion of edge pixels and therefore result in more mixed pixels.
Solution:
Get Kristin Abraham’s stories in your inbox
Join Medium for free to get updates from this writer.
In short: money and technology.
Within just the last decade, we have witnessed an explosion of very-high resolution sensors coming to market. The price tag still is a problem to apply this solution on scale. Often these images have to be tasked, which can lead to delays. If the research is limited to small geographic contexts, drones can be an interesting solution, too. Likely, within the next decade we will find that the trend to higher resolution will also be followed in public sensor launches by national agencies. The finer the resolution gets, the more “pure” pixels and the less “mixed” pixels are observed.
An alternative way to increase spatial resolution is through modelling. With advancements in Machine Learning and image correction, super-resolution algorithms have emerged that are based on training data to model higher resolution. Other than data fusion these algorithms do not constantly need new high resolution images to pair with the lower resolution, making it cheaper and quicker. Super-resolved images can greatly assist scene interpretation, but are limited for applications that require absolute accuracy such tree counting.
I have described my experience using super-resolution for agroforestry in another blog post.
3. Agroforestry as an application
Problem:
The term agroforestry is generally used to describe management systems where trees and shrubs are integrated into agricultural activity. Monitoring this type of application is trickier than other applications such as monitoring large industrial-type cropland or reforestation, where one can expect only one thing on the surface.
The type of agroforestry where one is adding trees to tree crop commodities (orchards, like coffee and cocoa) is especially challenging, because crop and non-crop are very similar. It is nearly impossible to tell apart what is part of the commodity and what is not. Even on a higher level analysis, it is difficult to confirm agroforestry with Remote Sensing. You will find that in most land cover maps these systems are often confused for forests. The FAO has recently published a paper describing exactly this problem (Forest monitoring good practices: mapping tree crop commodities).
Limitation:
Tropical agroforestry as an application is hard to detect with Remote Sensing at field level, because it is so similar to actual forest. In industrial farming orchards are well ordered and object recognition can be achieved by identifying tree crowns and lines of planted trees, mostly in contrast to the bare soil they are planted on. In agroforestry objects become less distinguishable from above and in the early years of planting, many non-crop trees will be hidden under the canopy of existing vegetation.
While Remote Sensing can typically confirm that there is vegetation as opposed to non-vegetation, it cannot well differentiate between very similar types of vegetation. That is because optical Remote Sensing measures spectral behavior which will look very similar for many trees, even if we had sufficient spatial resolution.
Solution:
In short: role with the punches.
There are several ways one could address separating crop from non-crop trees, like object-based image analysis (OBIA), using hyperspectral or LiDAR data. Often machine learning is involved to find better variables to distinguish agroforestry systems from surrounding vegetation. All these techniques require a spatial resolution/ point density high enough to assure the assessment of individual trees. And also temporal resolution can be critical to exercises such as identifying plant phenology.
LiDAR is the most adapt to capture structural tree insights such as height and shape. But it largely depends on which type of LiDAR you are using. There is currently only one space-borne sensor (GEDI) that with a beam size of 25 m and a point spacing of 60 x 600 m does not provide enough detail for the application. Airborne LiDAR is therefore way more interesting, but costly and only provides measurements for one moment in time.
Quite frankly, all these solutions can get quite complicated and expensive and I would question whether it really is necessary for your application.
The holistic approach is to accept that there is going to be a mix of vegetation present in a field and pixel rather than trying to painstakingly separate each object. Partly agriculture, partly new trees, partly old trees. Rather than monitoring individual trees, shift to monitoring the health of the entire plot and search for applications that allow for abstractions and have a lower need for accuracy, e.g. measuring trends rather than total amounts.
Conclusion
Each of these topics in themselves are already a challenge. But combined, this case is VERY challenging. As soon as one of the challenges is addressed by working outside of the ‘standard’ case, the application becomes easier. E.g. larger plot sizes, drier climate.
The one magic ingredient
However, one of the biggest assets to solving the puzzle I left to the end. Its not miracle models, it’s field data (aka ground truth), that is necessary to feed, calibrate and validate any good model. Because eventually you will end up including one or two models into your product. Nothing is better than actual observation from the ground. And I often have to laugh, when companies claim they can be more accurate with satellite data. No, they cannot. If they are good, they can be almost as accurate as ground data.
It is hard to come by consistent, consequently scaled field data and I have seen few products that use field data once the model is established. That form of validation is invaluable and anyone who is serious about their Remote Sensing product makes a point out of using field data, too.
Summary
Literature
Lechner, M.A.; Stein, A.; Jones, S. D. and Ferwerda, J.G. (2009): Remote sensing of small and linear features: Quantifying the effects of patch size and length, grid position and detectability on land cover mapping. Remote Sensing of Environment 113, 2194–2204, https://doi.org/10.1016/j.rse.2009.06.002.
Vajsová, B.; Fasbender, D.; Wirnhardt, C.; Lemajic, S. and Devos, W. (2020): Assessing Spatial Limits of Sentinel-2 Data on Arable Crops in the Context of Checks by Monitoring. Remote Sensing, 12(14), 2195, https://doi.org/10.3390/rs12142195.
