Do Simplified Models Pick the Wrong Rural Towers?

Planning a rural broadband network comes down to a few expensive choices: where to put towers or access points, how to link them back to the Internet, and how to power them. The algorithms that make these choices are almost always tested only inside the simplified world they were designed for. In that world, radio coverage is a flat circle of fixed radius, and the cost of a site is the price of its hardware. Field deployments show both assumptions are badly wrong. Installation work such as wiring, mounting and fiber routing was over a third of capital cost in an outdoor 5G testbed. A rural over-water link lost more than half its capacity to 20 dB tidal fades. Off-grid power for a single mast varied more than fourfold in cost depending on how it was designed.

This proposal asks a narrow, answerable question. When a planner optimized under simplified assumptions is re-scored under realistic terrain, realistic costs and measured signal variability, how much worse is its plan than the plan a realistic planner would have chosen? We also ask which simplification causes most of the loss. The output is a measured "decision regret" for existing rural planning methods, checked against signal measurements of real towers. It is not another planner. If simple models turn out to choose nearly the same sites, that negative result is worth publishing, because it tells the field where extra modeling effort is wasted.

Core Claim

We claim that the errors in rural planning models matter in two different ways. They matter for what a plan is predicted to cost and cover, which can be badly off. They also matter for which plan gets chosen, which may or may not be robust. Prior work does not separate these two effects.

We define the regret of a design $D_0$, chosen at a low-fidelity level, as

$$R(D_0) = \frac{C_3(D_0) - C_3(D_3^\*)}{C_3(D_3^\*)},$$

where $C_3$ is the total cost of a design under the highest-fidelity evaluator. $C_3$ includes hardware, site-specific installation and power cost, plus a penalty for population left below a service threshold. $D_3^\*$ is the best design found when optimizing directly at that level.

Our hypothesis is that regret is small for flat, sparse terrain. We expect it to be large where hills block signals and where grid and road access varies across candidate sites. We also expect most of the regret to come from the cost model rather than the propagation model. That second prediction is the one most likely to be wrong, and the experiments are designed to check it.

Grounding in Prior Work

Two existing planners are the methods we audit. Network Graph Generation Through Adaptive Clustering and Infection Dynamics places access points in Kilimambogo, Kenya, using real population maps (Facebook HDX) and real tower coordinates (OpenCellID). It treats each access point as a fixed 750 m disk and arrives at 39 access points over 400 km². Minimum Cost Configuration of Relay and Channel Infrastructure solves exact relay placement as a mixed-integer program on a flat 144 km² plane, needing about 22 hours with a modified Benders decomposition. Both papers report results only under their own assumptions. We keep their algorithms and change only the world used to score them.

Three studies supply the realistic ingredients and motivate the gap. Campus5G measured the installation share of capital cost and validated ray-traced coverage against an 11,000-sample walk test. Its best interpolation error was an RMSE of 7.25 dB. Tegola measured 20 dB temporal fading on a 19 km link. It also showed that a wind-solar hybrid could power a 23 W mast for £900, against more than £4,000 for solar alone. Universal Scanning GUI Tool for TVWS combines the Longley-Rice terrain propagation model with GMTED2010 elevation data at national scale. We reuse that propagation pipeline.

Two further papers constrain the design. Exploring High Tower Base Stations with Multi-User Massive MIMO projects 37 km cells from simulation alone, and its authors concede that phone uplink power is unverified. We therefore make uplink feasibility part of the highest-fidelity evaluator instead of assuming the downlink footprint. Lessons From the Commercial Failure of Project Loon reports break-even monthly fees of $2.50 to $7.50. We use those figures to convert cost per subscriber into an affordability check.

Research Program

The core artifact is a fidelity ladder of evaluators that share a single candidate-site set and population raster.

At level 0, each site covers a 2D disk and costs the same fixed amount. Level 1 replaces the disks with terrain-aware path loss from Longley-Rice over GMTED2010 elevation data, and checks both downlink and phone-limited uplink. Level 2 adds site-specific cost. That cost covers installation as a function of distance to roads and to the power grid, using Campus5G's measured installation share as a floor. It also covers off-grid power sized from local solar and wind climate data, following the method Tegola used. Level 3 adds signal variability: spatial shadowing with spread calibrated to measurement, and temporal fading drawn from measured link traces. A site then counts as serving a user only if the link holds for a target fraction of time.

Each planner is run at each level it can accept. Every resulting design is then scored at level 3. Ablations remove one ingredient at a time, which attributes regret to propagation, cost, or variability.

The empirical anchor is what keeps level 3 from being just another unvalidated model. Existing towers listed in OpenCellID act as free transmitters. We will collect commodity-phone drive-test measurements of signal strength from existing towers in a rural study area. These measurements test the level 0 and level 1 coverage predictions directly. They also give the shadowing spread used at level 3, instead of borrowing the 7.25 dB figure from an urban campus.

We will not have the Kenyan drive-test data before the campaign. Until then we will run the analysis with a range of spreads bracketing the Campus5G value. We will report results with that limitation stated plainly.

Baselines

The naive reference point is a uniform grid of identical sites, scored at level 3. The only number on this comparison is from Resident Population Density-Inspired Deployment of K-tier Aerial Cellular Network. It reports a coverage probability of 0.3845 for uniform placement against 0.9713 for population-matched placement with the same node count, in a stochastic-geometry arena. We expect the grid to be clearly beaten and include it only to set the scale of regret.

The main audited baseline is the two-phase clustering planner of Network Graph Generation Through Adaptive Clustering, run on its own Kilimambogo inputs. Its reported operating point is 39 access points at a 750 m radius, with up to a 35% backhaul length reduction from aerial nodes. We first reproduce that operating point. We then score it at level 3 under the same candidate sites and population raster as every other method.

The strongest competitor is the exact mixed-integer approach of Minimum Cost Configuration of Relay and Channel Infrastructure, which reports convergence in about 22 hours on its 144 km² instance. For it, we feed level 2 site costs and level 1 link feasibility into the same formulation. This "realistic exact planner" is the best design available without modeling variability. Every method gets the same 24-hour solver budget on the same machine.

A level 3 optimizer runs the same solver on scenario-sampled coverage. If its plan beats the realistic exact planner's plan by more than the run-to-run spread across random seeds, variability modeling earns its place. If not, level 2 is sufficient.

We found no reported result on Lacuna for regret of any rural planner under a change of modeling fidelity. All regret numbers will therefore be new.

Why not just…?

Why not just shrink the disk radius by a safety margin? This is the most serious threat. A disk calibrated to a fixed fade margin and fit to drive-test data may choose nearly the same sites as the terrain model. We include this calibrated disk as an explicit arm and fit its radius to the drive-test data. If its level 3 regret is within seed noise of the terrain-aware planner's, we will report that simple calibration suffices for propagation. That result would still leave the cost-model question open.

Why not just plug Longley-Rice into the planner and skip the audit? The TVWS tool already shows that terrain-aware propagation runs at national scale. Using it is not the open question. What is unknown is how wrong the terrain model itself is in rural terrain against phone measurements. It is also unknown whether propagation or site cost dominates the loss. The ablations and the existing-tower drive test answer both.

Why not just buy satellite terminals? Exploiting Hybrid Terrestrial/LEO Satellite Systems finds that satellite throughput peaks at roughly 200 to 400 satellites and that terrestrial sites serve dense clusters better. The Loon analysis shows that affordability, not coverage, is the binding constraint. We add a satellite-only arm, priced per subscriber, to the affordability comparison, so this trade-off is measured rather than assumed.

Proposed Experiments

The primary study area is Kilimambogo, chosen because its inputs already exist and the clustering planner's published result can be reproduced there. A second instance is a rural region of Ethiopia around a candidate high tower, which tests whether the super-cell plan survives the uplink constraint.

We report regret, coverage shortfall, cost per covered person compared against the Loon break-even fees, and rank agreement of designs across levels using Kendall's tau. The drive-test check compares predicted against measured signal strength from existing towers, as RMSE in dB, for disk, calibrated-disk and terrain models.

Three outcomes would be decisive. Regret above 15% for the level 0 plans would confirm that simplified planners choose wrong sites. Regret below 5% combined with large cost misprediction would mean simple planners rank designs correctly but quote the wrong prices. Calibrated-disk regret matching terrain-aware regret would mean terrain modeling can be dropped from the planner.