One district is shaped by drought, the other by floods. SoMAP is currently running a two-track assignment in Hobyo and Barawe that pairs a 720-household baseline survey with district-scale climate vulnerability mapping, so that targeting decisions rest on household evidence and spatial evidence together.

Hobyo District in Mudug Region, Galmudug State, and Barawe District in Lower Shabelle Region, South West State, are both climate-vulnerable, but not in the same way. Hobyo is an arid coastal and inland pastoral system where the binding constraints are rangeland condition, the distribution and spacing of water points, sand movement and the mobility of herds. Barawe is a riverine and coastal agro-pastoral system where the constraints are floodplain exposure, riverbank erosion, drainage and the extent and condition of cultivated land. Treating the two as variants of one problem would produce an average that describes neither.

SoMAP is undertaking an assignment in both districts built around that difference. Two tracks run in parallel under a single team lead. The first is a measurement exercise: a baseline survey producing defensible, reproducible values for outcome and impact indicators against a comparison group. The second is a decision-support exercise: a district and cluster-scale spatial diagnostic that narrows a long list of candidate community clusters to a defensible final set. The two have different sampling logics and different failure modes, so our team treats the combination as an efficiency measure rather than a methodological merger.

Measuring households, not impressions

Because no household listing exists for either district, the survey frame is being constructed in the field. The design is stratified, multi-stage and probability-based at every stage, so that every eligible household has a known and non-zero probability of selection and design weights can be computed rather than assumed. Four livelihood strata are applied inside each district, covering pastoral, agro-pastoral, riverine or coastal, and internally displaced and displacement-hosting households, which keeps the two districts comparable while allowing the stratum weights to differ.

Settlements are being selected with probability proportional to size, with cluster population estimates validated with district authorities beforehand. The design covers 720 households, 360 in the project arm and 360 in a comparison arm, across 36 clusters of 20 households selected systematically from a complete field listing. At 95 per cent confidence and 80 per cent power, and with an assumed intra-cluster correlation of 0.05, it carries a minimum detectable effect of roughly 0.29 standard deviations. That assumption is stated openly: the correlation will be re-estimated from the realised data so that later rounds can be sized on evidence rather than on an indicative figure.

The questionnaire runs across ten modules, covering demographics and assets, livelihoods, natural resource access and condition, experience of drought and flood shocks, participation in resource management institutions, service access, food security and social cohesion. Two design choices matter more than they appear. The women's module is administered by a female enumerator to an adult female household member rather than to the male household head on her behalf, and youth data come from a youth respondent inside the household rather than being inferred from the head's answers. Surveys that skip both under-record the access and decision-making dimensions the assignment is meant to track. A qualitative strand of 12 focus group discussions and 10 key informant interviews, purposively sampled within the same strata, explains the quantitative pattern rather than sitting beside it, and reports divergence between the two where it occurs instead of smoothing it away.

Quality control built into the instrument

All fieldwork uses electronic data collection with GPS capture on tablets, and quality assurance is arranged in four sequential layers. Controls that software can enforce are enforced by the instrument itself, through skip logic, range checks and mandatory fields. High-frequency checks are scripted before fieldwork begins and run nightly on the cumulative dataset, so that an enumerator effect surfaces in the first week rather than at analysis. District field leads accompany a share of interviews, back-check a further share by revisit or telephone, and approve each day's caseload after a debrief. Every record is validated against the assignment's own settlement layer, so a GPS point falling outside its assigned cluster is flagged the same evening.

Ten enumerators are being recruited locally, resident in their districts and trained together in a single session, and they carry a standard neutrality script. That script matters: no survey activity communicates eligibility for any benefit, which protects both data quality and the community relationships implementation will later depend on.

Spatial evidence at two resolutions

The diagnostic track began with a documented data inventory recording every dataset's source, year, coverage, resolution, reliability and limitations, so that a theme with no reliable source for one district is flagged rather than estimated. The analysis itself draws on a deliberately district-specific stack. Land cover is being classified from Sentinel-2 composites using a random forest classifier in Google Earth Engine, with accuracy reported through a confusion matrix, overall accuracy and kappa, and strengthened by GPS-tagged ground-reference points collected by the survey teams. Sentinel-1 radar carries the flood analysis in Barawe, because seasonal cloud degrades optical imagery there at exactly the time of year that matters. A hydrologically conditioned 12.5 metre ALOS PALSAR elevation model supports watershed delineation, slope, aspect and a flood depth proxy, alongside mapping of rivers, wadis and seasonal drainage. Drought is characterised from CHIRPS rainfall through a Standardized Precipitation Index and rainfall anomaly series, combined with MODIS vegetation condition anomalies. Boreholes, shallow wells and berkads are compiled from existing inventories with every attribute tagged as verified, indicative, incomplete or requiring field confirmation.

Those layers feed a vulnerability framework that treats vulnerability as exposure and sensitivity moderated by adaptive capacity. The sensitivity and adaptive-capacity components draw on household evidence from the survey track, so the vulnerability surface is not built on remote sensing alone. Two commitments accompany it. The weights used in the composite overlay are agreed in advance and stated on the face of every output, because an unstated weighting scheme makes an index unauditable. And the resulting ranking is tested against alternative weighting schemes, because if a ranking flips under plausible alternative weights the decision-maker needs to know before using it.

What the outputs will let people do

The work will deliver separate drought and flood exposure maps for each district, watershed and terrain layers, a water point and proximity analysis, land cover maps with stated accuracy, and cluster-level map sets at a resolution where individual gullies, dune fronts and riverbank erosion scars become visible. Everything comes in editable form: ArcGIS project, map and layer packages and a file geodatabase, Google Earth-compatible files sized for a tablet in the field, print-ready map sheets carrying a data source statement and a confidence note, and ISO 19115-compliant metadata for every layer. Processing is scripted throughout, so any layer can be regenerated from raw inputs and the method re-run for new districts without being re-procured.

What this demonstrates about SoMAP

This is the assignment type that defines our practice, because it requires survey research and geospatial analysis to be strong in the same organisation at the same time. The work is still underway and no findings are yet available. What the design already shows is how we approach evidence: probability sampling rather than convenience, stated assumptions rather than silent ones, confidence flags on every layer, reproducible code as a deliverable, and a willingness to report that an indicator cannot be reliably derived rather than fill the cell.