Satellite fire inventories miss most small agricultural burns. In Northeast China, the MODIS 500 m burned-area product (MCD64A1) catches only 22.5% of cropland burn scars that a 30 m Landsat analysis finds (Mapping Cropland Burned Area in Northeastern China). The same blind spot plausibly explains why aerosol models fed with standard inventories underestimate smoke by factors of 5 to over 10 in regions dominated by dispersed burning (Petrenko et al.). The usual explanation is "sub-pixel dilution": a burned field fills only a small fraction of a coarse pixel, so its signal is averaged away and the pixel is labelled "not burned." Spatial transcriptomics has the same problem. A Visium capture spot pools RNA from several cells, and the field has built methods that estimate what fraction of each spot belongs to each cell type. The strongest of these combine a reference library of pure signatures with a prior that neighbouring spots have similar compositions (Li et al. deconvolution benchmark).
This note proposes treating each coarse MODIS pixel the way a Visium spot is treated. Instead of a yes/no burn flag, we estimate a continuous burned fraction. The pure "burned" and "unburned" reflectance signatures come from sparse 30 m Landsat scenes. A spatial-autocorrelation prior (a statistical assumption that nearby pixels tend to have similar values) is restricted to neighbouring cropland pixels. The claim to test is that much of the missed burning is recoverable from data we already have. The spatial prior should be the ingredient that makes recovery work, and especially the ingredient that lets a model calibrated in one region work in another.
Core Claim
The 77.5% omission in MCD64A1 is largely a mixture problem, not a pure detection failure. Burning practice spreads across neighbouring fields within days, so burned fraction is spatially clustered. A reference-based unmixing model with a conditional autoregressive prior (a Gaussian prior whose precision matrix couples each pixel to its spatial neighbours) should recover per-pixel burned fraction at 500 m. The resulting county-level burned area should be far closer to the 30 m Landsat totals than MCD64A1, and the model should transfer across regions and years better than a supervised regressor trained on paired labels.
The specific transfer is the CARD-style formulation that the deconvolution benchmark found among the most robust. CARD is a reference-based deconvolution method with a spatial autocorrelation prior. The file gives no numeric accuracy values for that benchmark; it reports qualitative rankings only. Its satellite analogue, as far as the file shows, has not been attempted.
Research Program
For each MODIS pixel $i$ and burning window, let $y_i$ be the pre-minus-post change in reflectance-based burn indices (the Burned Area Index and Normalized Burn Ratio). We model it as
$$y_i = f_i\, s_{\text{burn}} + (1-f_i)\, s_{\text{crop}} + \epsilon_i,$$
where $f_i \in [0,1]$ is the burned fraction, $s_{\text{burn}}$ and $s_{\text{crop}}$ are reference change signatures estimated from Landsat pixels labelled burned or unburned, and $\epsilon_i$ is noise. The fractions get a conditional autoregressive prior:
$$\mathrm{logit}(f) \sim \mathcal{N}\!\left(\mu,\ \tau^{-1}(D - \rho W)^{-1}\right).$$
Here $W$ is a neighbour-weight matrix, $D$ is the diagonal matrix of its row sums, $\rho$ sets the strength of spatial coupling, and $\tau$ sets the overall precision.
Two design choices take the over-smoothing tension seriously. DUSTED needed attention weighting to avoid blurring cortical layer boundaries (Zhu et al.). The 3×3 majority vote in the Landsat study can erase isolated small burns. So $W$ links only cropland-to-cropland neighbours under the 30 m GLC_FCS30 mask, with weights decaying in spectral dissimilarity. In addition, $\rho$ is estimated rather than fixed, so the data can turn the prior off where burning is genuinely isolated.
Direction 2 enters in two further ways. Global and local Moran's $I$ of the model residuals serve as a diagnostic: well-specified unmixing should leave spatially random residuals, while leftover High-High residual clusters flag systematic omission, for example under smoldering or cloud. All validation also uses spatially blocked cross-validation, because random pixel splits leak autocorrelated signal and inflate scores.
The study region and year are Northeast China, 2015, where 39 Landsat 8 scenes and the multi-harmonic 30 m burn map exist. Transfer tests use other NEC sub-regions, and other years where Landsat maps can be regenerated with the same published method. The 30 m map is itself imperfect (Producer's Accuracy 84.2%, User's Accuracy 95.9%), so it serves as a reference standard of known error rather than as truth. Results are reported with that ceiling stated.
Baselines
The naive reference point is MCD64A1 as distributed, aggregated to 500 m burned fraction and to county totals. It is the product inventories actually use. Against the Landsat reference it reports Overall Accuracy 67.3% and Producer's Accuracy 22.5% (Mapping Cropland Burned Area). The 30 m GLAB AM product is a second off-the-shelf reference, with OA 68.5% and PA 20.4% in the same paper. Beating these is necessary but easy, so they only anchor the scale.
The strongest existing competitor is a supervised downscaling regressor: a gradient-boosted model that predicts the Landsat-derived burned fraction from the same MODIS change features plus 3×3 and 5×5 neighbourhood summaries. Neighbourhood summaries give it spatial context without a formal prior. No paper in the file reports this baseline on this task; no reported result found on Lacuna for supervised MODIS-to-Landsat burned-fraction regression. It gets the same inputs, the same cropland mask, the same spatial blocks, and a hyperparameter search matched in wall-clock budget to the tuning of $\rho$, $\tau$ and the weight kernel in our model.
The key ablation is identical unmixing with $\rho = 0$, a per-pixel linear unmixing with the same reference signatures. It isolates whether the spatial prior is what helps. A second ablation replaces the cropland-restricted, similarity-weighted $W$ with plain queen contiguity, to test the edge-preservation choice.
Beating a baseline means lower county-level burned-area bias and lower per-pixel RMSE of burned fraction on held-out spatial blocks. Under cross-region and cross-year transfer, the improvement must exceed the spread across blocks. At a matched threshold, Producer's Accuracy should rise well above 22.5% without User's Accuracy collapsing.
Why not just…?
Why not just use Landsat or Sentinel-2 everywhere? Where clear 30 m scenes exist, they are better; the Landsat method reaches 92.9% OA. Landsat's 16-day revisit, however, lets ploughing and regrowth erase scars between overpasses, and cloud removes further dates (Mapping Cropland Burned Area). The proposal uses Landsat as the reference library and the near-daily MODIS record as the data to unmix. We will measure how much of the 30 m signal survives when only one or two Landsat dates per season are available for building signatures. If the full 30 m record is routinely available in a region, this method is unnecessary there, and we will say so.
Why not just per-pixel linear spectral unmixing, which remote sensing already does? This is exactly the $\rho=0$ ablation. Evidence from tissue suggests spatial priors matter for low-abundance components. There, reference-based methods with spatial constraints outperformed alternatives, and reference-free factorization failed on rare constituents (Li et al. benchmark). A burned fraction of a few percent is such a rare constituent. If the ablation matches the full model, the borrowing from transcriptomics adds nothing and the claim fails.
Why not just train a supervised regressor with neighbourhood features? This is the real threat. Within one region it may match or beat our model, because it learns nonlinear structure directly from labels. Our bet is narrower: a model with physical reference signatures and an estimated spatial coupling should degrade less under geographic domain shift. Mobility-flow models show that kind of shift (Quantifying geographic domain shift). The cross-region and cross-year transfer experiment settles it. If the regressor transfers as well, we will report that spatial context matters but a formal prior does not.
Why not just a 3×3 majority filter or a GNN like STELLAR? A majority vote operates on binary labels, so it cannot create sub-threshold fractions, and it erases isolated burns. STELLAR's graph convolutions gave an 11% accuracy gain on cell annotation (Brbic et al.). That is a labelling problem with abundant annotated nodes, though, not mixture recovery under scarce labels. We include a small graph network trained on the same labels as an extra comparison, but we expect it to behave like the supervised regressor.
Proposed Experiments
The first experiment is in-region recovery. On NEC 2015, with spatially blocked folds, we compare per-pixel burned-fraction RMSE, county burned-area bias, and PA/UA across MCD64A1, the supervised regressor, the $\rho=0$ ablation, the contiguity-$W$ ablation, and the full model.
The second experiment, the decisive check, is transfer. We fit on one NEC sub-region or year and test on others, re-estimating only the reference signatures from the target's sparse Landsat dates. The proposal's central claim stands or falls on whether the full model's degradation under transfer is smaller than the regressor's.
The third experiment is diagnostic. We map local Moran's $I$ clusters of residuals and check whether they coincide with known failure modes such as cloudy windows and smoldering-prone fields (Deng et al.), separating residual omission from model misfit.
The fourth experiment is a downstream sanity check in Daqing. Using the same 50 km buffers and the upwind analysis, we ask whether corrected burned area correlates with ground AQI and PM2.5 at least as well as MCD64A1. The reported values for MCD64A1 are R = 0.82 for AQI, 0.88 upwind, and R² = 0.81 for PM2.5 (Comparative Analysis). Because those correlations are already high, this test can only show that the correction does not inject noise. The real gain should appear as a larger total burned area at similar correlation. That would be consistent with the AeroCom underestimation, and it motivates a later emissions-model test outside this note's scope.
References
- Mapping Cropland Burned Area in Northeastern China by Integrating Landsat Time Series and Multi-Harmonic Model
- Biomass burning emission analysis based on MODIS aerosol optical depth and AeroCom multi-model simulations: implications for model constraints and emission inventories
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- DUSTED: Dual-Attention Enhanced Spatial Transcriptomics Denoiser
- Quantifying geographic domain shift to decouple the geospatial transferability of human mobility flow generation models
- Annotation of spatially resolved single-cell data with STELLAR
- Source apportionment of black carbon aerosols from light absorption observation and source-oriented modeling: an implication in a coastal city in China
- Comparative Analysis of the Impact of Two Common Residue Burning Parameters on Urban Air Quality Indicators