| Variable | PRISM Combo | Bias (°F/in) | MAE | RMSE | r | MAE Bar | Notes |
|---|
| Variable | Best Combo | MAE | 2nd Best | MAE | Spread |
|---|---|---|---|---|---|
| Tmin | 800m OFF | 2.762 | 800m ON | 2.771 | 0.009 |
| Tmean | 4km OFF | 3.114 | 800m OFF | 3.119 | 0.005 |
| Tmax | 800m ON | 5.050 | 800m OFF | 5.052 | 0.002 |
| Tdmean | 800m ON | 3.410 | 800m OFF | 3.412 | 0.002 |
| Ppt | 4km ON | 0.1109 | 800m ON | 0.1109 | 0.0000 |
| Station | N days | Tmin bias | Tmin MAE | Tmean bias | Tmean MAE | Tmax bias | Tmax MAE | Ppt bias | Ppt MAE | QC flag |
|---|
| Station | N | 800m ON bias / MAE | 800m OFF bias / MAE | 4km ON bias / MAE | 4km OFF bias / MAE | Best | ||||
|---|---|---|---|---|---|---|---|---|---|---|
SLVN01 (Sullivan Co): 135 obs removed by QC. Remaining 686 days still show tmin bias −4.85°F vs PRISM and RMSE 19.96°F — the largest error in the network even after cleaning. Tmin values reaching −110°F suggest a systematic sensor failure during cold/low-light conditions (likely thermistor or ADC issue at low voltages). PRISM shows −9°F as its minimum; observed minimum hit −110°F. This station's tmin should be flagged as UNRELIABLE until hardware inspection. Tmax and ppt appear more reasonable.
CLNT01 (Clinton Co): tmin RMSE 10.57°F vs 800m-ON — second-worst in network. 15 bad obs removed. High variance relative to PRISM suggests either (a) valley cold-pool microclimate that PRISM cannot resolve at 800m, (b) sensor siting in a frost pocket, or (c) intermittent sensor noise. Bias is near-zero (+0.55°F tmax) so it is not a systematic offset — it is genuine variance. This is actually a scientifically interesting station if the high values are real cold-pool events.
Spring Tmax MAE = 7.0°F (network-wide): The largest seasonal error across all stations and combos. Likely driven by PEMN sensors receiving direct solar radiation heating before full canopy closure (March–May), causing inflated observed tmax relative to PRISM's radiatively shielded ASOS-anchored surface. PRISM is correct; PEMN stations may need radiation shielding audit for spring period. Consider applying a spring-specific correction to PEMN tmax in this period before using as training data.
| Use Case | Recommended Combo | Reason |
|---|---|---|
| RFSI training feature (all temperature vars) | 800m ON | Marginally best for tmax/tdmean. Uses PRISM's full DEM-aware bilinear blend — physically appropriate for point-location training |
| Tmin specifically | 800m OFF | Very slight edge (0.009°F MAE) — negligible in practice. Either 800m combo is fine |
| Precipitation (all uses) | 800m ON or any | All 4 combos produce identical results. PRISM's precip surface is smooth enough that resolution and interpolation make no measurable difference |
| Bias characterization / delta field | 800m ON minus 4km ON | Resolution delta is the signal — shows lapse-rate impact of DEM cell mismatch per station. Use as metadata covariate in RF |
| Grid-level prior (background field) | 4km ON | 4km aligns with ERA5 downscaling resolution; marginally better tmean. Use as background, 800m as correction surface |
| Station-to-grid cell audit | 800m OFF | "Which exact PRISM cell is this station in?" — OFF gives raw cell value, no blending ambiguity |