In field experimenting with satellite imagery in agriculture. What’s holding us back?

This is my perspective drawn my own experience as a grain farmer and having built tools for satellite imagery in the past. My journey started out as research into drone imagery in 2013 and led to a clear focus on the applications of the freely available Landsat 8 and later Sentinel 2 multispectral satellite data.


When it comes to how satellite imagery is applied to agriculture, it seems that up until most recently, many that provided satellite imagery usually offered just one, at most five spectral indices, with by far the most common being Normalised Difference Vegetation Index (NDVI). Clearly this is starting to change with more bandwidths available and the work going into machine learning that aims to provide some additional insights. NDVI is definitely the low hanging fruit when it comes to the types data that is available (medium spatial resolution (10 to 30m), multispectral, medium width bands) as it does a good job of estimating vegetation properties such as leaf area index (LAI) or more broadly crop health.

The fact that NDVI is so good at giving a broad spatial indicator of crop health leads many to believe that NDVI is in fact all there is to satellite imagery. The End. But the issue here is that an enormously complex procedure with unlimited options for processing and visualizing the data captured has been reduced to a somewhat opinionated idea that a particular algorithm and color scale is what you need. This is GREAT when it comes to selling a ‘solution’ to a busy agronomist. The idea being let the remote sensing experts figure out what the agronomists need to do their job and sell it to them. It’s a solid option.

The alternative

I’d like to think about an alternative where the agronomist is empowered to ‘do the work’ or at least have the option to when they need to dig deeper. Let’s say you combined an experienced agronomist or farmer with the power to test out 5 different spectral indexes over a field. They may discover that EVI or MTVI2 deals better with the impacts of wet soil on a cotton field that had just been irrigated before the latest image acquisition or that specific leaf and stem responses of herbicide drift are most noticeable in one algorithm over another. They are continuously on the ground and able to compare what the analysis tells them versus what reality is.

Sure, these results are not going to be peer reviewed ‘science’ but they are definitely impactful and can drive formal research into that direction or some may want to build up some IP around investing in better understanding lesser known spectral indexes to offer better solutions to their clients.

Going further, the red edge bandwidths in Sentinel 2 have allowed for nitrogen estimation algorithms. It could be that slow uptake with these is because some of the solutions built and sold are black boxes that don’t allow for self-directed experimentation. Again, it makes sense that an innovator comes in, learns about the technology, and sells the ‘solution’ to the agronomist but discussions I’ve had with some agronomists is that they would like to have more control over that box or at least be able to look inside.

Risks and obstacles

One risk is an agronomist making a decision based on a pattern they think they are seeing or some other misguided insight that is not there. I’d argue that this happens with standard NDVI and that encouraging more experimentation and looking in the box will lead to better informed agronomists. Therefore, the obstacle here is the agronomists current technical understanding of satellite imagery data and process. A further obstacle is time to allocate to these types of experiments and upskilling instead of just always defaulting to NDVI. But as remote sensing becomes more prevalent in agriculture a well-informed agronomist should probably have the skillset required to understand this process. Also, time invested in understanding satellite imagery can usually be offset by human productivity gains.

Another factor hindering exploration and experimentation may be the focus on AI/machine learning as the solution. Despite the potential benefits, this emphasis may overshadow the vast opportunities for discovery and innovation through field experimentation by skilled agronomists using the full range of spectral indices.

Ultimately, the extent to which a farmer or agronomist uses satellite imagery is their choice. However, it is hoped that this information will encourage them to consider the available options more thoroughly.

Did you know there are 100+ spectral indices available (Awesome Spectral Indices).