💻 computer science

Intercomparison of Machine Learning Algorithms for Remote Sensing-based In-season Crop Mapping

This study evaluates ten machine learning algorithms using Harmonized Landsat-Sentinel imagery and crop rotation data to demonstrate that Support Vector Machines can accurately map corn in Iowa and almonds in California at 30m resolution by early June, providing a near real-time tool for in-season crop monitoring and emergency response.

Original authors: August Posch, Jitendra Kumar, Forrest M. Hoffman, Auroop R. Ganguly

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: August Posch, Jitendra Kumar, Forrest M. Hoffman, Auroop R. Ganguly

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a farmer or a food safety official trying to guess what crops are growing in a field right now, in the middle of the growing season. Usually, the government (USDA) only tells you what was planted after the harvest is over, often months later. But if a flood or a pest outbreak happens now, you need to know immediately.

This paper is like a massive "cooking competition" where ten different computer chefs (machine learning algorithms) try to figure out what's growing in the fields just by looking at photos taken from space. The goal is to see which chef can guess the crop type most accurately by early June, before the harvest.

Here is a simple breakdown of what they did and what they found:

The Ingredients (Data)

The chefs were given two main ingredients to work with:

Space Photos: They used a special mix of satellite photos (from NASA and European satellites) that take pictures of the Earth every few days. They looked at how the plants reflect light (like how a red shirt looks different from a green shirt) over time.
The Family History: They also gave the computers a "family tree" of the land. They looked at what was planted in that exact spot for the last four years. For example, if a spot grew almonds for four years in a row, it's very likely to grow almonds again.

The Contestants (The Algorithms)

The researchers tested ten different types of computer brains (algorithms). Some were like simple rule-followers, while others were complex pattern-recognizers. They tested them on two very different challenges:

The Iowa Corn Challenge: In Iowa, the fields are huge and mostly filled with corn and soybeans. The problem here is that corn and soybeans look very similar to each other, like trying to tell apart two identical twins.
The California Almond Challenge: In California, the farms are smaller and mixed with many different crops (walnuts, grapes, tomatoes). The problem here is that almonds are rare in the mix, like finding a specific needle in a haystack full of other needles.

The Rules of the Game

To make it a fair test, they didn't just let the computers practice on last year's data and then guess on next year's data. Instead, they used a "Yearly Cross-Test."

They trained the computers on data from four years (e.g., 2018–2021).
Then they asked the computers to guess the crop for the fifth year (2022), which they had never seen before.
They repeated this five times, swapping which year was the "test" year. This is crucial because weather changes every year, and a model that works in a wet year might fail in a dry one.

The Results: Who Won?

After running thousands of different settings (like adjusting the seasoning on a dish), they found:

The Best Chefs: The winners were Support Vector Machines (SVMs). Think of these as the most precise chefs who can draw very sharp lines between different crops.
- For California Almonds, the best model got a score of 0.74 (out of 1.0). This means it was pretty good at spotting almonds early in the season.
- For Iowa Corn, the best model got a score of 0.59. This was lower because corn and soybeans are so similar that even the best computer struggled to tell them apart early on.
Early vs. Late: Surprisingly, the models that only saw photos up to early June performed just as well as models that saw the entire year's photos. It turns out, the "family history" (what grew there before) combined with just the first few months of photos was enough to make a good guess. You don't need to wait until the end of the season.
The "Sample Size" Surprise: The researchers thought they needed to feed the computers every single pixel of data (millions of points). But they found that feeding them just 0.1% of the data (a tiny sample) worked just as well as feeding them 100%. It's like tasting a spoonful of soup to know if it needs salt; you don't need to eat the whole pot to know the flavor.

The Catch: The Weather is Unpredictable

The biggest finding wasn't just about which algorithm won, but how much the results changed from year to year.

In some years, the models were great. In others, they struggled.
Why? Because farming changes. In 2020, for example, the amount of almond land in California suddenly tripled. The computer, trained on old patterns, got confused because the "family history" didn't match the new reality.
This means there is no single "perfect" setting that works for every year. The best settings for a wet year might be wrong for a dry year.

The Bottom Line

This study proves that we can build computer models to map crops while they are still growing, which is vital for reacting to disasters like floods or pests. The best tool for the job is a Support Vector Machine, but we have to accept that the accuracy will wiggle up and down depending on the weather and farming decisions of that specific year.

The authors suggest that in the future, we might need to combine these models with extra information (like weather forecasts or economic news) to help the computers handle those surprise years better.

Technical Summary: Intercomparison of Machine Learning Algorithms for Remote Sensing-based In-season Crop Mapping

Problem Statement
In-season crop mapping is essential for food security, particularly as climate change increases the frequency of extreme weather events threatening agricultural yields. While the USDA Cropland Data Layer (CDL) provides high-resolution (30m) crop type maps, it is only available post-harvest (typically the following February), rendering it useless for near-real-time emergency management or yield forecasting during the growing season. Existing in-season mapping studies often lack rigorous evaluation across multiple unseen years or fail to systematically compare a wide range of algorithms under standardized conditions. Consequently, there is a gap in understanding which machine learning (ML) algorithms generalize best to unseen years, accounting for interannual variability in weather, phenology, and farming decisions.

Methodology
The study conducted a systematic intercomparison of ten machine learning algorithms to map two specific crop types: corn in Iowa and almonds in California. The research utilized the following data and processing pipeline:

Data Sources: Harmonized Landsat-Sentinel (HLS) surface reflectance time series (30m resolution) and historical CDL data for crop rotation history.
Study Areas: Two 100km x 100km regions: Sacramento, CA (almonds) and Ames, IA (corn).
Feature Engineering: HLS data was filtered for clouds, shadows, and snow, then temporally composited into two schemes: 5-day and 14-day periods. Features included spectral reflectance (Blue, Green, Red, NIR, SWIR1, SWIR2) and Boolean crop rotation history (presence of the target crop in the prior four years).
Algorithms: Ten algorithms were evaluated: Random Forest (RF), Extra Trees (ET), Logistic Regression (LR), Linear Support Vector Machine (LS), Nonlinear Support Vector Machine (SV), Gaussian Naïve Bayes (NB), K Nearest Neighbors (KN), Neural Network Multilayer Perceptron (NP), Principal Components K Nearest Neighbors (PK), and Principal Components Logistic Regression (PL).
Hyperparameter Optimization: A grid search was performed for each algorithm, resulting in 791 unique hyperparameter configurations.
Validation Scheme: A robust year-wise 5-fold cross-validation was employed using data from 2018–2022. In each fold, models were trained on four years and validated on the remaining unseen year. This approach quantified uncertainty due to interannual variability.
Metrics: Performance was evaluated using Precision, Recall, and the primary metric, F1 score.
Scale: Approximately 9,492 model configurations were run. Initial screening used 0.1% random samples of the data; top configurations were re-evaluated using 100% of the data via a bagging approach to confirm results.

Key Contributions

Systematic Intercomparison: The study provides a fair, standardized comparison of ten diverse ML algorithms across identical datasets, preprocessing steps, and validation protocols, addressing the lack of comparability in prior literature.
Robust Uncertainty Quantification: By validating across five unseen years, the study quantifies the range of performance expected in operational settings, moving beyond single-year evaluations that may not capture interannual variability.
Hyperparameter-Phenology Linkage: The research investigates how specific hyperparameter settings interact with interannual variations in crop phenology and distribution, offering insights into why certain models succeed in specific years.

Results

Algorithm Performance: Support Vector Machines (both Linear and Nonlinear) emerged as the most successful algorithms overall.
- California Almonds (Early June): Logistic Regression and Linear SVM achieved the highest mean F1 scores of 0.74 (standard error ~0.15). Nonlinear SVM, K Nearest Neighbors, and Random Forest also performed well (F1 > 0.71).
- Iowa Corn (Early June): Linear and Nonlinear SVMs tied for the best performance with a mean F1 score of 0.59. Performance was generally lower than for almonds, attributed to the phenotypic similarity between corn and soybeans and frequent crop rotation between the two.
Early-Season vs. End-of-Season: Models restricted to data available only by early June achieved performance within 0.01 of models trained on full-year data. This suggests that early-season reflectance combined with crop rotation history is sufficient for accurate mapping.
Training Data Size: Training on 0.1% of the available data yielded performance statistically indistinguishable from training on 100% of the data, indicating that spatially representative sampling is sufficient for these tasks.
Interannual Variability: Significant variation in performance was observed across validation years. For example, the best almond model's F1 score ranged from 0.47 to 0.88 across the five years. In 2020, a year with a tripling of almond acreage, models struggled due to a lack of historical continuity in crop rotation patterns.
Hyperparameter Sensitivity: Optimal hyperparameters varied significantly by year. For instance, in K Nearest Neighbors, the best distance metric (Manhattan vs. Chebyshev) and neighbor count ( $K$ ) depended heavily on the specific crop rotation history profile of that year.

Significance and Claims
The paper claims that its primary significance lies in providing practitioners (farmers, emergency managers, policymakers) with a robust, uncertainty-quantified expectation of in-season crop mapping performance. By demonstrating that Support Vector Machines are the most reliable algorithm across diverse conditions, the study offers a concrete recommendation for future operational systems. The authors emphasize that while no single model configuration is perfect for every year due to phenological and distributional shifts, the identified best practices (e.g., using SVMs, leveraging crop rotation history, and accepting early-season data) provide a strong foundation for near-real-time decision-making. The study concludes that future improvements may be achieved through ensemble approaches and the integration of ancillary data (e.g., weather, economic drivers) to anticipate years with unusual crop distributions.