Computer Science > Computer Vision and Pattern Recognition
[Submitted on 16 Sep 2026]
Title:AgroBench: A Reproducible Multimodal Benchmark for Weakly Supervised Crop Yield Learning from County Statistics and Pixel Observations
View PDF HTML (experimental)Abstract:Reliable agricultural yield statistics are typically reported at coarse administrative scales, whereas modern geospatial machine learning methods require spatially explicit, pixel level supervision. This mismatch has limited the development of large-scale benchmarks for crop yield learning using multimodal Earth observation data. A reproducible benchmark, AgroBench, is presented for transforming publicly available U.S. county level crop yield statistics into weakly supervised pixel-level crop time series. Each crop pixel time series is paired with a county-level yield value as a weak supervisory signal rather than a directly measured pixel-level yield label. Our geospatial data generation pipeline integrates USDA crop yield statistics with crop-specific land cover masks, Sentinel 2 multispectral imagery, Sentinel-1 synthetic aperture radar observations, climatic variables, and terrain information to produce temporally aligned multimodal sequences describing individual crop pixels throughout the growing season. The resulting benchmark contains over 13 million observations from 788,654 unique crop pixels spanning 5,107 county year combinations across eight growing seasons (2017 to 2024) for five major U.S. crops. To facilitate standardized evaluation, we establish a crop yield prediction benchmark using a Leave-One-Year-Out evaluation protocol and provide baseline results using representative machine learning models. By releasing the complete data generation pipeline, benchmark dataset, and evaluation protocol, AgroBench provides a reproducible foundation for future research in weakly supervised learning, multimodal remote sensing, spatiotemporal modeling, and geospatial foundation models for agriculture.
References & Citations
Loading...
Bibliographic and Citation Tools
Bibliographic Explorer (What is the Explorer?)
Connected Papers (What is Connected Papers?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)
Code, Data and Media Associated with this Article
alphaXiv (What is alphaXiv?)
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Hugging Face (What is Huggingface?)
ScienceCast (What is ScienceCast?)
Demos
Recommenders and Search Tools
Influence Flower (What are Influence Flowers?)
CORE Recommender (What is CORE?)
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Facts Only
* AgroBench is a multimodal benchmark for weakly supervised crop yield learning.
* The dataset transforms U.S. county-level crop yield statistics into pixel-level crop time series.
* Data sources include USDA crop yield statistics and crop-specific land cover masks.
* Imagery used consists of Sentinel-2 multispectral and Sentinel-1 synthetic aperture radar observations.
* Additional variables include climatic data and terrain information.
* The benchmark contains over 13 million observations.
* Data covers 788,654 unique crop pixels.
* The scope includes 5,107 county-year combinations across eight growing seasons (2017 to 2024).
* The dataset focuses on five major U.S. crops.
* Evaluation is conducted using a Leave-One-Year-Out protocol.
Executive Summary
AgroBench addresses a significant data mismatch in agricultural machine learning: the gap between coarse, county-level yield reporting and the pixel-level resolution required for geospatial modeling. By pairing individual crop pixel time series with aggregate county yield values, the framework enables weakly supervised learning, bypassing the need for direct pixel-level yield labels which are rarely available at scale.
The pipeline integrates a diverse array of multimodal Earth observation data, combining optical and radar imagery with climate and terrain variables. Spanning eight growing seasons and five major U.S. crops, the resulting benchmark provides a standardized foundation for developing spatiotemporal models and geospatial foundation models. The inclusion of a Leave-One-Year-Out evaluation protocol and baseline results ensures that future research can be measured against a consistent and reproducible standard.
Full Take
This work operates in ACADEMIC MODE. The methodology attempts to solve the "label scarcity" problem in remote sensing by leveraging weak supervision. A peer reviewer would likely focus on the inherent noise introduced by using county-level aggregates as proxies for pixel-level performance. While the sample size is massive (13 million observations), the fundamental confound is the assumption that county-level statistics provide a meaningful supervisory signal for a specific pixel, despite significant intra-county variance in soil quality, farm management, and micro-climates.
The claim of "reproducibility" is well-supported by the release of the full data generation pipeline, which extends existing knowledge by integrating SAR (Sentinel-1) and optical (Sentinel-2) data into a unified temporal sequence. However, the utility of the benchmark depends on whether the models can actually extract signal from the weak labels or if they simply converge on the county average.
For this to matter beyond the lab, the models must demonstrate that they can generalize across different geographies where USDA-style reporting might not exist. The bridge to a higher level of validity would be a validation study using a small set of "ground truth" pixel-level yield measurements (e.g., from precise harvest monitors) to see how closely the weakly supervised predictions align with actual field reality.
The next logical step is to test this framework on global datasets where administrative reporting is even coarser or less reliable than in the U.S.
Counterstrike Scan: A coordinated campaign would use this as "proof" that AI can now perfectly predict food security to justify removing human oversight in agricultural policy. The actual content is a technical benchmark for ML researchers and does not match this pattern.