Keeping Local Crop Damage Visible in Cheap Spatial Yield Forecasts

Forecasting a county's corn yield in June or July, months before harvest, gives governments, insurers and traders time to act. Neighboring fields share weather and soils, so models that look at neighbors are a natural choice. The usual tool is a graph neural network (GNN), a model that repeatedly averages each location's features with those of its neighbors. That averaging has two costs. It is expensive once the graph is large. It is also a smoothing operation, and smoothing is harmful when the most important signal is a sharp local drop. During the August 2020 Iowa derecho, a windstorm that flattened crops along a narrow track, a spatial smoothing method treated the real damage as noise and erased it (Unsupervised Graph Spectral Feature Denoising for Crop Yield Prediction).

We propose to drop learned message passing and compute fixed neighborhood statistics once. The graph links satellite pixels, and the statistics include not only neighborhood means but also contrasts ("how much worse is this pixel than its surroundings") and dispersion ("how uneven is this neighborhood"). These statistics are pooled from pixels up to counties and fed to a small tabular model. The claim to be tested has two parts. First, this setup matches learned spatio-temporal GNNs on early-season county yield at a small fraction of their compute. Second, it does clearly better in years with localized shocks, because the contrast features carry exactly the information that smoothing destroys. Working at pixel level is also where scalability becomes a real problem rather than a nominal one.

Core Claim

The spatial GNNs used so far in yield prediction work on county graphs of roughly one hundred to one thousand nodes. At that size, speeding up the GNN barely matters. Fiedler-based edge pruning, which removes edges while preserving how well the graph stays connected, cut training on 99 Iowa counties from 52.70 s to 15.01 s (Graph Sparsification for GCN Towards Optimal Crop Yield Predictions). That is a real ratio, but the absolute time was already small.

The real bottleneck is elsewhere. Damage from hail, wind, flooding or localized drought happens at sub-county scale. A model that wants to see it needs a graph over hundreds of thousands to millions of cropland pixels. At that scale a learned GNN runs into neighborhood explosion: each extra layer multiplies the set of pixels a prediction depends on, until the whole graph must sit in GPU memory. Evidence that neighborhood aggregation need not be learned comes from Fixed Aggregation Features Can Rival GNNs. That paper converts graphs into tables of multi-hop mean, sum, max, min and standard-deviation features, feeds them to a multilayer perceptron, and matches or beats learned GNNs on 12 of 14 node-classification benchmarks. Its evidence is on citation and social graphs for classification. We carry the idea to a regression setting with physical meaning, change what is aggregated so that it targets discontinuities, and add a pooling step from pixels to counties.

Concretely, let $x$ be a vector holding one GDD-aligned feature for every cropland pixel. (GDD, growing degree days, is accumulated heat; aligning dates by GDD rather than the calendar lines up crop development stages across regions.) Example features are an EVI composite or land surface temperature at a given stage. Let $P$ be the row-normalized sparse adjacency of a pixel graph that links nearby pixels with similar soils. For hops $k=1,\dots,K$ we compute

$$m_k = P^k x,\qquad c_k = x - m_k,\qquad s_k = \sqrt{P^k (x\odot x) - m_k\odot m_k},$$

where $m_k$ is the $k$-hop neighborhood mean, $c_k$ is the pixel's contrast against that mean, $s_k$ is the neighborhood's local spread, and $\odot$ is elementwise multiplication. Each county's features are quantiles (5th, 25th, 50th, 95th) and the fraction of pixels with $c_k$ below fixed thresholds, taken over that county's cropland pixels. We also compute the same statistics for the anomaly relative to the pixel's own multi-year mean. Computing everything takes $O(K|E|)$ sparse multiplications, where $|E|$ is the number of graph edges. It runs on a CPU, needs no backpropagation, and must be redone for each new season only up to the current date.

Research Program

The work has three parts.

The first is building the graph and the features. Pixels are MODIS 500 m cells masked to corn using the USDA Cropland Data Layer, as in the YieldNet data. Edges link each pixel to its 8 spatial neighbors and to a small number of soil-similar pixels within a radius, using soil attributes (clay, available water storage, organic carbon) as in the spectral denoising work. Time series are resampled by GDD, following Early Season Forecasting of Corn Yield at Field Level from Multi-Source Satellite Time Series Data, which reports MAPE falling from 17.02% to 16.25% under thermal-time alignment.

The second is the downstream model. We use gradient-boosted trees and a small MLP, each fed the pooled county features together with county soil and weather covariates. Both are cheap to train and run.

The third is a scaling study. We measure memory and wall-clock time of feature construction against the number of pixels, from one state up to the 13-state Corn Belt. We compare this with a learned GNN trained on the same pixel graph using neighbor sampling. We also compare with a pixel-level GNN run with coarsened inference in the style of FIT-GNN, which partitions large graphs and summarizes boundary context to cut inference memory. This part of the study makes the compute claim measurable rather than asserted.

Baselines

The naive reference is a per-county linear trend fitted to historical yields. It uses no satellite data and sets the floor any forecast must clear.

The non-spatial baseline is XGBoost on county-mean remote sensing, weather and soil features. This follows the multi-source tree ensembles shown by Liangliang Zhang et al. to outperform single-source inputs.

The strongest distributional baseline is YieldNet (Simultaneous corn and soybean yield prediction from remote sensing data using deep transfer learning). It compresses each county's pixels into histograms over time, reflectance and band, then applies a CNN. It reports corn MAE of 8.74% of average yield on the 13-state Corn Belt, and 7.8% at one month before harvest. It is the strongest competitor because its histograms already capture how pixel values are spread within a county. It is run on our exact county-years and splits, trained under the same hyperparameter search budget.

The strongest spatial baselines are two. The first is EvolveGCN-H on a Fiedler-sparsified county graph (Graph Sparsification for GCN Towards Optimal Crop Yield Predictions); no reported MSE value was found on Lacuna for this model, so we rerun it. The second is XGBoost on EVI denoised with the graph Laplacian regularizer, a spatial smoothing penalty (Unsupervised Graph Spectral Feature Denoising for Crop Yield Prediction), which reports $R^2 = 0.5929$ on the 2020 test year across 938 counties in 10 states. That same year and region is our shock-year test, so this number is the direct target.

Every model receives identical input variables, the same leave-one-year-out splits, and an equal tuning budget, measured in trials. Beating the field means lower RMSE than YieldNet and the best GNN on held-out years, and $R^2$ above 0.5929 on the 2020 10-state test at the same June–July cutoff.

Four ablations isolate what matters. The first keeps only the mean features $m_k$, which is plain smoothing. The second adds $c_k$ and $s_k$. The third computes the same statistics on a county graph rather than a pixel graph. The fourth trains a learned GNN on the identical pixel features, which tells us whether learning the aggregation adds anything.

Why not just…?

Why not just use a sparsified learned GNN, as Bagheri et al. did? Their result is on 99 counties of a single, fairly uniform state. Its message passing is built from averaging, which the derecho case suggests is the wrong inductive bias for localized shocks. The "learned GNN on the same pixel features" ablation settles the question directly. If it beats the fixed features in shock years at acceptable cost, our claim fails.

Why not just use county-level tabular features with no spatial structure? Tree ensembles on multi-source data are strong, and on normal years they may be close to everything else. The test is concentrated error: we report RMSE separately for counties inside the 2020 derecho footprint. The prediction is that county means dilute a narrow damage track, and contrast fractions do not.

Why not just use YieldNet-style histograms? This is the most serious threat. A within-county histogram already records the share of low-EVI pixels. The difference is that a histogram cannot tell a damaged pixel from a pixel that is always low, such as poor soil. Contrast against neighbors and against the pixel's own history can make that distinction. We add YieldNet histograms as extra inputs to the non-spatial tree baseline. If that model matches ours in the derecho footprint, then the spatial structure is unnecessary, and we will report that result.

Why not just scale a GNN with neighbor sampling or coarsened inference? That approach is feasible. The question is whether its cost buys any accuracy. The scaling study reports peak memory, GPU hours and CPU inference latency for each approach side by side.

Proposed Experiments

The main experiment uses the 10-state and 13-state Corn Belt county yields from USDA-NASS, 2004–2020, with MODIS EVI, NDVI and LST masked by the Cropland Data Layer. Forecasts are issued at GDD-aligned cutoffs corresponding to mid-June, mid-July and mid-August, with leave-one-year-out evaluation. The Iowa 99-county setup allows a direct rerun of the Fiedler-sparsified GNN.

We report RMSE, MAE as a percentage of average yield, and $R^2$. We also report error broken down by yield decile, since regression toward the mean in high-yield regions is a known failure of these models, for example STF-MoE (Dynamic gating-enhanced deep learning model with multi-source remote sensing synergy for optimizing wheat yield estimation). For compute we report feature construction time, training time and peak memory.

The decisive check is 2020 in the derecho footprint. The approach succeeds if contrast and dispersion features cut error there relative to both the means-only ablation and the Laplacian-denoising pipeline, while staying within noise of the best method on normal years. It must do this at a CPU-only cost at least an order of magnitude below a sampled pixel-level GNN. If histogram-augmented trees close the gap, the conclusion changes: within-county distributions suffice, and pixel graphs are not needed for early-season county forecasting.