From Roadside Surveys to Satellite-Based Maps: Mapping Tillage Practices in Eastern Washington

By Amin Norouzi Kandelati, Department of Biological Systems Engineering, WSU

With support from a CSANR BIOAg grant, Amin Norouzi Kandelati developed a machine-learning framework combining satellite imagery, roadside field observations, and crop-group information to improve tillage mapping across eastern Washington.

How farmers till their fields—from intensive soil disturbance to no-till management—can affect erosion and soil carbon. Knowing where different practices are used can help researchers and conservation organizations track regional trends and, eventually, estimate their effects across the landscape. But eastern Washington contains far more fields than researchers could realistically inspect in person. Satellite-based tillage mapping can help fill that gap by combining observations from a sample of fields with satellite imagery and machine-learning models.

When I started looking into mapping tillage practices across eastern Washington, I thought the path would be relatively straightforward. Collect some ground data, apply existing satellite-based approaches developed in the Midwest, and produce tillage maps for the region.

It turned out not to be that easy.

Many existing approaches use satellite imagery to estimate the amount of crop residue left on a field and relate that residue to tillage intensity. That logic works reasonably well when more intensive tillage generally means less residue. But in eastern Washington, that relationship is more complicated.

We have tremendous diversity in crops, climate, and topography. More importantly for our work, different crops leave very different amounts of residue. Legumes, for example, can leave relatively little residue even under no-till management because they have low biomass to begin with. From a satellite image, a no-till legume field can look surprisingly similar to a conventionally tilled wheat field.

That was the first lesson of this project: models developed for one region will not always work perfectly in another. While this is intuitive, what surprised me was just how poorly they generalized here, and how dramatically performance degraded for some crop groups.

Inside the tillage mapping project

Satellite imagery is only part of the story. This short video follows WSU researchers and conservation professionals as they combine field observations, agricultural expertise, and machine learning to map tillage practices across eastern Washington.

Learning from the roadside

Joel Demory from Washington State Department of Agriculture teaches Amin how to identify crops. Photo: Kirti Rajagopalan, WSU

We also faced a more basic problem. Unlike some regions where tillage observations have been collected for years, we were largely starting from scratch.

Even collecting the ground data was more complicated than I expected. Identifying tillage from the roadside is not always clear-cut. Residue provides clues, but so do crop type, field conditions, and other signs of disturbance. Ultimately, there is some judgment involved.

I spent a lot of time in the field with Extension experts, particularly Steve van Vleet, learning how to use a combination of clues to assess tillage. I also gained an appreciation for what a “ground-truth observation” actually represents. Behind each point in our dataset was time in the field and, often, expert judgment.

When existing models didn’t work

Once we had enough observations, we could finally test the approaches developed for other regions.

They didn’t work particularly well.

Performance was especially poor for some cropping systems. For legumes, for example, the baseline approach correctly classified tillage class only about 34% of the time.

So instead of simply applying an existing model, we had to rethink the problem. We developed a more general modeling framework that explicitly considered crop type and accounted for the fact that the relationship between residue and tillage differs among cropping systems.

That approach made a substantial difference. Across our 577 field observations, median overall accuracy increased from 69% using the baseline approach to 84% using the framework we developed. For legumes—the crop group that had given us the most trouble—accuracy increased from 34% to 76%.

The improvement in accuracy allowed us to move from hundreds of roadside observations to tillage predictions across thousands of fields in eastern Washington. The full methods and results are available in our preprint, Regional-scale Field-level Estimation and Mapping of Tillage Practices in Areas with Crop Diversity.

And that is where the project became much more interesting to me.

From a research map to something people could use

Producing a map is one thing. Producing a map that someone actually finds useful is another.

I worked with stakeholders, including Ryan Boylan, Bradly Johnson, and Put linked [1] after the paragraph reporting the 69%–84% and 34%–76% results., to learn how conservation professionals might use tillage maps. These conversations led to several rounds of evaluation and revision.

Three people look together at a screen with maps, while one points at the maps. — Bradly Johnson, Ryan Boylan, and Amin discuss changes in tillage practices over time. Photo: Darrell Kilgore, WSU

The maps were then used to look at regional trends in conservation tillage adoption, including work supporting the Voluntary Stewardship Program. Putting the maps to use steered the project in a new direction. The question was no longer simply, “Can we map tillage?” It became, “What does someone need to know before they can actually use this map?”

One answer was uncertainty.

An overall accuracy number tells a stakeholder how well the model performs as a whole, but it doesn’t describe how much confidence to place in the prediction for a particular field. Stakeholders wanted to know where predictions were more reliable and where they should be treated with greater caution.

So uncertainty became the next phase of the project. We began developing methods of quantifying uncertainty for each field-level prediction rather than treating uncertainty as something summarized only for the model as a whole.

That led to another realization: uncertainty could also tell us something about where we should collect data next.

Our original field observations were collected using fairly traditional geographic sampling approaches. But if collecting ground observations is expensive, why collect the next observations in places where the model is already confident? Interestingly, we found that utilizing uncertainty itself, together with diversity in crops and growing conditions, to iteratively collect new data could result in matching the full-data performance with fewer data, suggesting that this approach could guide future data collection.

In other words, the model could help us improve not only the tillage map, but also help inform where to collect the data used to build the next map. We describe this approach in more detail in our preprint on uncertainty-guided data collection.

Looking beyond crop type

Accounting for crop type helped improve predictions using satellite imagery considerably, but it also raised another question. Crop type affects how much residue is left behind, but what about the age and degradation of that residue? Crop type is not typically accounted for in large-scale machine-learning applications using satellite imagery, yet smaller-scale experiments suggest it could matter. As residue decomposes and weathers, its characteristics change, potentially affecting what a sensor sees and, in turn, how we interpret tillage.

To explore this concept, I worked with Haly Neely and colleagues on a spectroscopy experiment measuring the spectral reflectance of different crop residues at different stages of degradation. We found that residue age significantly affected the spectral signal and, consequently, tillage predictions.

This experiment was smaller and more controlled than our regional satellite mapping, but it demonstrated something important that large-scale models tend to simplify: the residue we see in a field is not static. The type of residue matters, but what has happened to that residue since harvest can be equally important. A related preprint on residue type, residue age and soil characteristics examines how these factors can bias satellite estimates.

When big data is small

The tillage mapping work helped expose a broader challenge for agricultural AI: Satellite imagery can generate enormous datasets, but the field observations needed to train and evaluate models remain scarce and expensive. Kirti Rajagopalan considers what that disconnect means—and what researchers can do about it—in her related Perspectives piece, Agricultural Data: Big or Small?

Read Kirti's perspective

Where we go next

The project continues to evolve, and there is still plenty to do.

One direction I am particularly excited about is moving beyond mapping tillage. Tillage affects erosion and soil carbon, so we plan to use these spatially explicit tillage maps as inputs to erosion and soil-carbon models. This approach will allow us to ask not only where tillage practices are changing, but also what impacts various tillage practices have on soils across the landscape.

Amin, advisor Kirti Rajagopalan, and Palouse Conservation District partner Ryan Boylan learn about the costs and benefits of no-till cropping from Zac. Photo: Darrel Kilgore, WSU

We are also working to extend this new approach to mapping using satellite imagery beyond eastern Washington. Something as simple as field boundaries becomes a challenge when moving into areas where good field-boundary datasets are not readily available, so we are developing ways to create boundaries and expand the work to other dryland agricultural areas of the Pacific Northwest.

Looking back, the project has turned out a little differently from what I imagined when I started. I thought the hard part would be applying satellite imagery and machine learning at scale. Instead, some of the most important lessons came from figuring out what the satellite images couldn’t tell us, learning from people who understand these cropping systems, and talking with the people who might actually use the maps we produce.

And perhaps that insight is what has been most interesting about this work: the better the technology becomes, the more it points us back to the people who will use it and the agricultural systems it represents.

What piqued your interest reading this? Please let me know.

Amin Norouzi Kandelati is a PhD candidate in Biological Systems Engineering advised by Kirti Rajagopalan. This blog post is based on parts of his PhD dissertation work, some of which was funded by the BIOAg program, and some in collaboration with the AgAID Institute. Contact Amin at a.norouzikandelati@wsu.edu.

Footnotes

[1] Details can be found in Norouzi Kandelati, A., Stahl, A. T., Yan, Y., Chaudhary, S., Van Vleet, S., Gharsallaoui, M. A., Doppa, JR., Gustafson, D., Kok, H., & Rajagopalan, K. Regional-scale field-level estimation and mapping of tillage practices in areas with crop diversity. Under review. Preprint available at 10.2139/ssrn.5035781.

[2] Norouzi Kandelati, A., Gharsallaoui, M. A., Dubey, N., Doppa, JR. & Rajagopalan, K. Uncertainty-Aware Tillage Mapping in Data-Limited Agroecosystems. Under review. Preprint available at 10.2139/ssrn.7345028.

[3] Norouzi Kandelati, A., Neely, H., Chaudhary, S., Alicea, S., Stahl, A. T. & Rajagopalan, K. Uncovering the hidden errors: How residue type and age impact remote sensing of crop residue cover. Remote Sensing of Environment. Preprint available at 10.2139/ssrn.7345026.