Where to add new stations
Judging the stations that exist and choosing sites for new ones are two different problems. A station that scores poorly as an index site is not, for that reason, a good place to put another one. A new station is most useful where two things hold at once: the existing network constrains the snow field poorly there, and snow there contributes strongly to downstream flow. This is the maximum-information, minimum-redundancy principle of monitoring-network design (Li et al. 2012; Mishra and Coulibaly 2009), applied here geostatistically rather than through entropy.
The score is a heuristic priority, built in the spirit of information-based network design rather than derived from it. It multiplies two terms: how much new snow information a site would carry, measured as the posterior variance of a Gaussian process at that cell (Zimmerman 2006; Krause et al. 2008), and how strongly snow there loads onto runoff, measured as the interpolated hydro_B of the nearest stations that carry it. Under a linear-Gaussian model a gain of that shape factorises into a product, which is why the two multiply and neither needs a weight; the analogy is close enough to motivate the form and not close enough to make the product an estimate of forecast improvement. The kriged loading in particular is an interpolation of a measured quantity, not the coefficient that a formal value-of-information calculation would require. Raleigh et al. (2025) support the premise that where you observe matters more than how much; they do not validate this score. One declared parameter, β, sets how far the product is traded against terrain the network under-samples.
One elevation rule, stated as eligibility. A candidate cell must reach 1000 m; below that it is not a candidate. Above the floor nothing re-weights by elevation — a cell at 1100 m and a cell at 2600 m are judged on coverage, loading and novelty alone. It is a floor and not a ramp on purpose: a ramp would keep penalising by elevation inside the band it is meant to admit, and the rain–snow transition is precisely where a warming winter changes the hydrology first (Klos et al. 2014; Godsey et al. 2018, on eleven years of transition-zone observations at Johnston Draw in Reynolds Creek, Idaho). In a warm winter a high station can still record a healthy peak SWE while much of the basin below it is taking rain.
1000 m is a judgement, and here is the reason for it. With no floor at all the shortlist puts four sites between 445 and 955 m, in canyon and irrigated-valley country: the lower Clearwater near Lewiston, the Weiser valley, the Magic Valley. Their downstream loading is 0.18–0.30, among the lowest such a list produces, so they are chosen because no station is near them rather than because their snow drives any river. With respect to the objective this project chose — monitoring a seasonal snowpack — those cells are artefacts of a coverage term computed over ground the network was never built to watch; they are not artefacts merely because they are low. The floor removes that ground and costs 15 % of the state. It is an exploratory eligibility choice and not a mapped boundary of the rain–snow transition: that boundary varies with local climate, with the season and with the individual storm (Klos et al. 2014; Godsey et al. 2018), and nothing here locates it.
Fig 17.How the siting score is built. The coverage gap and the downstream loading multiply into a value of information. That product is divided by its 99th percentile over the state so it can be compared with terrain novelty, and β sets the balance between the two. Nothing else re-weights the score. A flat eligibility floor at 1000 m decides which cells are candidates at all. Sites are chosen one at a time, and the field is re-scored after each pick, so the second site is not simply the second-best cell of the first hotspot.
Three decisions the map depends on
The correlation length is fitted from the stations. The coverage gap is a posterior variance, and its one parameter is the distance over which snowpack anomalies stay alike. We estimate it from the station records: for every pair, the correlation of their 1-April SWE against their separation. Each year’s state-wide signal is removed first, and that step matters. Without it the pair correlations stay near 0.35 even at 700 km, which only reflects that a wet winter is wet across the whole state, and the fit returns 202 km. With the year signal removed the correlation crosses zero near 150 km and the fitted length is 50 km (1711 pairs, 59 stations).
The network does not stop at the state line. A posterior variance is always largest at the edge of the data it is given. Computed from Idaho stations alone, it reported a gap two kilometres from the Utah border, with a Utah SNOTEL twenty kilometres beyond it. The gap is therefore conditioned on 308 stations: the 87 in Idaho plus 221 within 150 km of the state line in Montana, Wyoming, Utah, Nevada, Oregon and Washington. Candidate cells still have to lie inside Idaho.
The grid has to be square. 3DEP is asked for 500 m but receives a bounding box in degrees, and for Idaho it returns a projected raster whose pixels are 371 m by 446 m. Slope computed on that grid, or on a reprojection of it to degrees, is wrong in one direction. Everything here is computed on a true square 500 m grid in EPSG:5070, 866,738 cells inside the state. Longitude and latitude are derived from that grid for the mask and the maps, and are never fed back into a calculation.
Fig 18.The three fields the score is made of, each rescaled to [0, 1] over Idaho, and — fourth panel — the floor on the candidate set: a cell must reach 1000 m to be eligible, and 15 % of Idaho does not. The floor is an eligibility rule, not a weight. Filled dots are Idaho SNOTEL, open dots the out-of-state stations the coverage gap is also conditioned on. The gap and the novelty fall to zero at an existing station by construction, which is the check that each field behaves as intended.
Fig 19.Left: the part of the score that follows from the data alone. Right: the full score, with the shortlist numbered in the order the greedy search chose it. The difference between the two panels is what the terrain-novelty term adds. The right panel is zero below the 1000 m candidate floor; above it nothing re-weights by elevation, so a cell in the rain–snow transition band competes on coverage, loading and novelty like any other.
What it returns
Of the 866,738 Idaho cells that carry an elevation, 737,016 reach the 1000 m floor and compete. The shortlist keeps its mountain core and adds a transition-zone band the old score could not reach. Five of the twelve sites are in the central Salmon River Mountains and the Frank Church country, between 44.6° and 45.2°N; two more are in the Lost River and Lemhi ranges and one in the southeastern highlands near Portneuf. The leading site is at 45.18°N, 115.25°W, 1785 m, where the coverage gap is 0.40 and the downstream loading 0.69. The other four are on the margins: 1162 m above the lower Clearwater, 1020 m on the north side of the Magic Valley, and two in the Owyhee uplands at 1237 and 1505 m.
By elevation band the twelve split six above 2000 m, three between 1500 and 2000 m and three between 1000 and 1500 m; the lowest sits at 1020 m and the mean at 1793 m. Run without any floor, the same score keeps ten of these twelve and puts four valley-bottom cells below 1000 m in place of the other two, so the floor is doing exactly the one job it is there for and nothing else.
β moves part of the shortlist, and that should be said plainly. Matched one to one, nine of twelve sites at β = 0.25 and ten of twelve at β = 0.75 fall within 25 km of a β = 0.50 site; the median matched separation is 0 km in both cases, so the sites that agree are the same cells, but two to three move a long way. By half-degree cell the overlap with β = 0.50 is Jaccard 0.64 and 0.46. The mountain core survives all three values of β; the picks between 1000 and 1500 m are the ones that move.
What it does not say
Terrain novelty does little to order the shortlist. Eight of the twelve sites sit at the top of its scale. Only 1.0 % of Idaho cells reach that level, so those eight really do lie in the most unusual one per cent of terrain; but with most candidates tied at 1.00, the term mainly separates eligible cells from ineligible ones and leaves the ordering to coverage × loading. The term is also terrain only. The two canopy predictors of the representativeness index have no grid-wide counterpart, so a cell whose vegetation is unlike anything the network samples is not flagged for that reason.
Half of Idaho lies beyond the loading measurements. hydro_B exists at 43 of the 87 stations. 51 % of grid cells lie further than one correlation length from any of them, and there the loading field is the network mean rather than a measurement, so a high score in those cells rests on coverage and novelty alone. The interpolation also returns negative values over 13.4 % of the state, although no station measures one. A loading is a weight on information and cannot be negative, so those cells are set to zero before the field is scaled, and the count is reported rather than absorbed into the rescaling.
A low site needs instrumentation that resolves the precipitation phase. Three picks sit between 1000 and 1500 m, where the snow pillow will read zero for part of the winter — and zero on its own does not distinguish rain from no precipitation from snow that has already melted. A standard SNOTEL already records SWE, precipitation and air temperature together, which resolves much of that; what these sites would need is that the three are read jointly, and ideally a sensor that identifies the phase directly.
Calling them transition-zone sites would be a claim we have not made. They are low relative to the rest of the shortlist, which is a statement about elevation, not about snow regime. Placing a cell in the rain–snow transition needs a climatic classification — snow fraction of precipitation, or an observed phase record — and this analysis contains none. The same caution applies to calling the valley-bottom cells the floor removes artefacts: they are artefacts with respect to the objective of monitoring a seasonal snowpack, which is the objective we chose, not because they are low.
The 1000 m floor is a judgement, and it is the only one of its kind in the score. It is an exploratory eligibility choice, not a mapped boundary: it excludes ground where a snow pillow would read zero for most of the winter, and it is not derived from any rain–snow classification. It is not fitted to anything and its sensitivity is not reported here. Moving it would move the low end of the shortlist; the mountain core would not notice.
No feasibility filter. Access, land status, wilderness designation and power are not considered, and the mountain picks fall in country where the first three would be the binding constraint. The map says where a station would be most informative; whether one can be built there is a separate question.
Recommendations
Representativeness is a question of station, scale and season, not a fixed label. The network-mean index falls from about 0.56 at 500 m to about 0.42 at 4 km, and the observed ratio departs most from 1 at onset and during melt.
Use the physiographic index as a first-order screen, not an automatic verdict. It is robust to the weights. Its best use is to rank stations and flag candidates for LiDAR or field follow-up. It has been compared with airborne LiDAR at one site, which is a check and not a validation of the state-wide ranking, so it is a screen rather than a measurement.
Judge stations on both axes. We detect no monotone relationship between physiographic fidelity and downstream information value, over an interval wide enough to leave a moderate association possible, so a station that looks unrepresentative may still be one of the most useful for forecasting. Stations that score low on both deserve review first.
Focus new monitoring on the central Salmon River Mountains and the Frank Church–Gospel Hump country, where the coverage gap and the downstream loading are high together, and secondarily on the Lost River and Lemhi ranges. Both are terrain the present network constrains poorly and whose runoff it predicts least well; neither recommendation survives a feasibility review unexamined.
Treat the rain–snow transition band as part of the network, not as ground outside it. Six of the twelve priority sites fall below 2000 m and three below 1500 m. These are the elevations a warming winter changes first, and the ones a score built to find persistent snow will never propose. They need a site that records precipitation and air temperature alongside SWE, because there the absence of snow is the measurement.
Next steps for the index itself: explicit wind and forest-gap process terms, a snow-state rather than calendar seasonal variable, naturalized flows for regulated reaches, and a canopy term in the siting novelty score, plus a feasibility and land-status filter for proposed sites.
Citations
The seventeen works cited in the text above, covering the point-to-area problem, physiographic controls on snow, scale dependence, LiDAR as a reference, and forecast value and monitoring-network design.
Currier, W. R., and Lundquist, J. D. (2018). Snow depth variability at the forest edge in multiple climates in the western United States. Water Resour. Res., 54, 8756–8773. doi:10.1029/2018WR022553
Deems, J. S., Fassnacht, S. R., and Elder, K. J. (2006). Fractal distribution of snow depth from lidar data. J. Hydrometeorol., 7, 285–297. doi:10.1175/JHM487.1
Garen, D. C. (1992). Improved techniques in regression-based streamflow volume forecasting. J. Water Resour. Plan. Manage., 118, 654–670.
Godsey, S. E., Marks, D., Kormos, P. R., Seyfried, M. S., Enslin, C. L., Winstral, A. H., McNamara, J. P., and Link, T. E. (2018). Eleven years of mountain weather, snow, soil moisture and streamflow data from the rain–snow transition zone – the Johnston Draw catchment, Reynolds Creek Experimental Watershed and Critical Zone Observatory, USA. Earth Syst. Sci. Data, 10, 1207–1216. doi:10.5194/essd-10-1207-2018
Grünewald, T., and Lehning, M. (2015). Are flat-field snow depth measurements representative? A comparison of selected index sites with areal snow depth measurements at the small catchment scale. Hydrol. Process., 29, 1717–1728. doi:10.1002/hyp.10295
Herbert, J. N., Raleigh, M. S., and Small, E. E. (2024). Reanalyzing the spatial representativeness of snow depth at automated monitoring stations using airborne lidar data. The Cryosphere, 18, 3495–3512. doi:10.5194/tc-18-3495-2024
Klos, P. Z., Link, T. E., and Abatzoglou, J. T. (2014). Extent of the rain–snow transition zone in the western U.S. under historic and projected climate. Geophys. Res. Lett., 41, 4560–4568. doi:10.1002/2014GL060500
Krause, A., Singh, A., and Guestrin, C. (2008). Near-optimal sensor placements in Gaussian processes: theory, efficient algorithms and empirical studies. J. Mach. Learn. Res., 9, 235–284.
Li, C., Singh, V. P., and Mishra, A. K. (2012). Entropy theory-based criterion for hydrometric network evaluation and design: maximum information, minimum redundancy. Water Resour. Res., 48, W05521. doi:10.1029/2011WR011251
Meromy, L., Molotch, N. P., Link, T. E., Fassnacht, S. R., and Rice, R. (2013). Subgrid variability of snow water equivalent at operational snow stations in the western USA. Hydrol. Process., 27, 2383–2400. doi:10.1002/hyp.9355
Mishra, A. K., and Coulibaly, P. (2009). Developments in hydrometric network design: a review. Rev. Geophys., 47, RG2001. doi:10.1029/2007RG000243
Molotch, N. P., and Bales, R. C. (2005). Scaling snow observations from the point to the grid element: implications for observation network design. Water Resour. Res., 41, W11421. doi:10.1029/2005WR004229
Molotch, N. P., and Bales, R. C. (2006). SNOTEL representativeness in the Rio Grande headwaters on the basis of physiographics and remotely sensed snow cover persistence. Hydrol. Process., 20, 723–739. doi:10.1002/hyp.6128
Raleigh, M. S., Small, E. E., Bair, E. H., Wobus, C., and Rittger, K. (2025). Snow monitoring at strategic locations improves water supply forecasting more than basin-wide mapping. Commun. Earth Environ., 6, 665. doi:10.1038/s43247-025-02660-z
Sexstone, G. A., Fassnacht, S. R., López-Moreno, J. I., and Hiemstra, C. A. (2022). Subgrid snow depth coefficient of variation spanning alpine to sub-alpine mountainous terrain. Cuad. Investig. Geogr., 48, 79–96. publisher page
Winstral, A., and Marks, D. (2014). Long-term snow distribution observations in a mountain catchment: assessing variability, time stability, and the representativeness of an index site. Water Resour. Res., 50, 293–305.
Zimmerman, D. L. (2006). Optimal network design for spatial prediction, covariance parameter estimation, and empirical prediction. Environmetrics, 17, 635–652. doi:10.1002/env.769