bluegreencity-ki

Irrigation Planning

Status

Produktiv seit 2026-09-24 (issue #227, Phase 3 Smoke-Lauf verifiziert: 89.280 Kandidaten, 84 ausgewählte Entwürfe, 3.395 L am Planungstag, Policy irrigation-policy-v0.1; Grafana-Dashboard „Bewässerungsbetrieb” liest die Decisions live). Der tägliche 04:00-Zyklus ruft den Planner ab enabled: true auf; die Policy-Versionierung und Kalibrierung (#223, v0.2) laufen parallel und ändern die live-Values ohne Rewrite.

Phase 2 provides a deterministic, UI-independent draft planner. It combines the latest Phase 1 soil-water state with the ML damage score and stable site vulnerability and allocates complete irrigation demands across a configurable multi-day resource horizon.

The result is a draft decision, not an automatically released work order. Taipy review, overrides, and approval belong to Phase 3.

Decision Logic

Candidates are eligible when either their combined risk reaches the configured threshold or their forecast SWBM stress reaches the urgent threshold. The planner excludes zero demand, irrigation avoidable by forecast rain (irrigation_avoidable_by_rain, see swbm_7_day_forecast.md section 7), and dates inside the minimum irrigation interval.

The canonical products remain separate:

• ml_damage_score_t1: provisional ML score for damage in year t+1; it is not presented as a calibrated probability;
• site_vulnerability_rulebased: stable site vulnerability from five equally weighted effect groups (water_soil, root_site, exposure, species_resilience, and chronic_load);
• swbm_forecast_stress: maximum modeled stress over the forecast horizon;
• irrigation_action_priority: an action-ranking score, not another risk estimate.

Action priority is a staged cascade, not a blend of three equally ranked models (design: internal planning note “Kaskade: Label & Feature-Reduktion”, 2026-08-30, Block A):

1. Gate (hard): a tree without modeled water demand (irrigation_demand_liters <= 0, i.e. SWBM layer-1 available water still above the 0.4 trigger from the original SWBM script) gets priority 0, regardless of ML score or site vulnerability.
2. Urgency (soft): among the trees that need water, the SWBM forecast stress (weight 0.5) is the dominant ranking signal.
3. Damage risk: the ML t+1 damage score (weight 0.3) refines the ranking.
4. Site importance: the stable site vulnerability (weight 0.2) refines it further.

Stages 2-4 are a weighted sum; stage 1 multiplies the result by the demand gate. This makes the SWBM the first-class selection criterion instead of one factor among three, and prevents a vulnerable site or high ML score from creating an irrigation action without a current modeled water deficit.

The SciPy MILP solver maximizes total selected priority under daily constraints for water volume, team minutes, vehicle-derived job count, trigger dates, and minimum irrigation intervals. It may therefore select several smaller jobs instead of one large higher-ranked job when their total priority is greater.

Current defaults for strategic value, heat vulnerability, service time, weights, and thresholds are transparent starting assumptions. They are not scientifically or operationally calibrated values.

Inputs

The planner only runs when irrigation_planning.enabled: true in config/config.yaml; the CLI and the daily cycle both refuse to plan while it is false (issue #227).

Input sources may be file paths (Parquet, GeoPackage, GeoJSON, CSV) or PostGIS schema.table references. The canonical production setup reads from PostGIS:

• bgcki.swbm_daily_state — SWBM demand, stress, trigger, rain, model, and quality fields; the read is anchored on the latest state_date on or before the planning date, so no local mirror is needed;
• bgcki.operational_tree_product — the daily-refreshed Lagebild table with stable objectid, ml_damage_score_t1, and site_vulnerability_rulebased.

Optional:

• canonical irrigation-event Parquet for minimum-interval enforcement;
• a weather forecast table with date and daily maximum temperature for heatwave detection;
• strategic value, heat vulnerability, and service-minute columns configured in irrigation_planning.columns.

Legacy inputs may still map drought_risk_ml to the ML component. Missing required priority products exclude the affected trees during input preparation and are counted in the CLI summary.

Phase 1 may emit several profile/forcing scenarios for one tree. The default conservative_max strategy creates one deterministic envelope from the latest day: maximum demand and stress, earliest trigger, minimum data quality, and rain avoidance only when every scenario agrees. Set the strategy to error to reject ambiguous scenario inputs instead.

The default stable-ID overlap gate requires at least 95% of latest SWBM trees to match the risk input. This deliberately rejects older pipeline artefacts whose objectid contains a year-coded inventory-record ID instead of stable Baum ID.

Capacity model (fleet calibration, 26.09.2026 — issues #227/#246)

The daily constraints come from irrigation_planning.resources. The current values are a raw estimate (Rohschätzung), not a calibrated value — the config comments carry the derivation: 9 fleet units (1× 10,000 L and 8× 3,000 L watering trucks, each operated by 1 driver) with max_jobs_per_vehicle: 60 jobs → 540 jobs/day. The shift switch selects the effective shift length from shift_hours (night = 8 h without traffic, day = 7 h with traffic), which scales the team-minute budget; daily_water_liters: 108000 is the raw daily water estimate. Reference measurement from the LHH: 62 trees/day/driver (LHH Kick-Off 07/2025); per-tree amounts come from the SWBM. Open LHH questions (tank counters, refill frequency, hydrant throughput, IBC reserve vehicles) are tracked in issue #246.

Run Locally

Set the real risk path and resource assumptions in config/config.yaml, or provide paths explicitly:

conda run -n bgcki_wesemann python scripts/run_irrigation_planning.py `
  --config config/config.yaml `
  --planning-date 2026-08-04 `
  --state-path data/derived/water_balance/daily_state.parquet `
  --risk-path data/output/<run>/hannover_gesamt/trees_at_risk_hannover_gesamt.gpkg `
  --no-postgis

--no-postgis writes only the configured Parquet history. Without it, apply sql/migrations/2026-08-04_create_irrigation_planning_tables.sql first; the same decisions are then upserted to bgcki.irrigation_plan_decision.

Outputs

One auditable record is written per eligible candidate and policy run:

• selection flag and recommended date,
• complete modeled amount in litres,
• normalized priority and level,
• explicit selection or exclusion reason,
• risk, current stress, forecast stress, and data quality,
• planning date and policy version.

The primary key (tree_id, planning_run_date, policy_version) makes reruns idempotent. Changing policy behavior requires a new policy version.

Validation

The deterministic planner was benchmarked with 61,193 synthetic candidates, seven daily capacity windows, and the configured 30-second solver limit. It selected 588 feasible assignments in 16.05 seconds. This validates the full-city computational scale, not the scientific policy values.

A real Phase 1 state file contains 91,585 profile/forcing scenarios for 61,193 stable tree IDs. The scenario envelope reduces these deterministically to one candidate per tree. The available 2026-08-02 risk GeoPackage cannot be used for an operational join: its objectid values are year-coded inventory-record IDs and overlap 0% with stable SWBM IDs. The 95% gate rejects it explicitly. A fresh risk pipeline output created after the stable-ID correction is required for the first real full-city planning run.

Literature Basis (Kalibrierung, s. Issue #223)

The policy values above are starting assumptions. The following sources are the literature basis for locally calibrating triggers, amounts, young-tree rules, and exclusions. Retrieved via Consensus, 22.09.2026.

Trigger thresholds (when to irrigate):

• Hahn et al. 2025, Ecological Indicators — ecophysiological irrigation thresholds for Central Europe: lower threshold = stem water reserve not refilled overnight (Tree Water Deficit), upper = 12 % xylem conductivity loss. Template for SWBM-stress trigger calibration.
• Schütt et al. 2022, Urban Forestry & Urban Greening — SWP monitoring at 17 young street-tree sites (Hamburg): critical availability at SWP < −1200 hPa, 3–5 months/year; critical zone migrates from root ball into the planting pit with tree age; age is a key predictor for critical water availability.
• Niu et al. 2025, Plants — stress thresholds are species-specific (stable at SWC ≈ 40 % vs. requiring > 60 %) — supports “no universal scheme”.
• Cheviron et al. 2024, J. Environmental Management — methodology for context-specific, locally calibrated irrigation rules with periodic reassessment; the methodological template for #223.
• Hahn et al. 2026, Ecological Informatics — low-cost thermal camera + CWSI thresholds for operational irrigation monitoring (future measurement points).
• Kluge et al. 2024, Env. Res. Communications — capillary rise delays water stress onset by ~45 days (49 Berlin lindens); confirms the mechanistic water-balance approach as a management tool.

Amounts and frequency:

• Rosenberger et al. 2025, Urban Forestry & Urban Greening (PDF in local Literatur und Quellen) — weekly manual irrigation does not prevent drought stress; daily rainwater-harvesting irrigation does; required storage is species-, site-, and climate-dependent.
• Haupt et al. 2025, Nature-Based Solutions — field trial in Hannover (Tilia cordata, weekly vs. monthly): irrigation raises cooling capacity by up to −0.67 °C crown temperature on >30 °C days.
• Torres et al. 2024, Frontiers in Climate — 25 % irrigation reduction has minimal productivity impact (transpiration linear in irrigation, LAI nonlinear) — basis for fraction-based allocation rules by priority class.
• Tams et al. 2024, Sci. Total Environ. — passive irrigation (system:catchment 1:1–1:2) substantially reduces but does not eliminate drought stress in young urban trees.
• Schütt et al. 2022, Forests — urban planting soils have very low plant available water capacity (6–10 % v/v) with species-specific growth deficits — site parameters must enter the amount rules.

Young-tree special rules:

• Schütt et al. 2022 (above, UFIG), Grey et al. 2018 Landscape & Urban Planning (underdrain doubles growth in 18 months), Thom et al. 2022 Water Resources Research (passive irrigation: 2× growth in year 1, 5.1 vs. 2.3 L/d transpiration; waterlogging caution in fine soils).

Exclusions (waterlogging / groundwater):

• Grey 2018 + Thom 2022: waterlogging is the central counterpart of drought — irrigate only where drainage/exfiltration exists. Planner hook: exclusion rule from the existing kriged groundwater surface in the SWBM.

Documentation and feedback loop:

• Vera et al. 2021 J. Hort. Sci. (sensor-based automation saves water, energy, labour vs. classical ET scheduling), Pascual et al. 2019 Sensors (sensor mesh for establishment monitoring), Cheviron 2024 (periodic rule reassessment) — support the “log date + amount in the tree register, feed back into the SWBM” requirement.

Agricultural context (method transfer only): Luketich et al. 2019 PLoS ONE, Marchionni et al. 2021 Landscape & Urban Planning, Young et al. 2026 Sustainability, Mohamadzade et al. 2021 UFIG, Korjenic 2025 Sensors, Waldburger et al. 2025 Agr. Water Manage. (dendrometer automation saves 38–45 % water), Domínguez-Niño et al. 2020 Agr. Water Manage.

Gap: no strong direct source for “watering deeper than 50 cm promotes deeper rooting” was found in the Consensus search; the local Vico et al. 2013 (Ecohydrology of street trees design and irrigation requirements) already covers this topic.

Zotero library round (22.09.2026, second round)

Additional sources from the BlueGreenCity-KI Zotero library. Zotero cite keys in parentheses; full metadata in Zotero.

Local calibration data (highest priority for #223):

• LHH soil-moisture sensor pilot (separate LHH project — not a BlueGreenCity-KI deliverable), based on LHH documents „Kommunale Klimafolgenanpassung” (2023) and „Einzelheiten zur Stadtbaumbewasserung” (2023). ~20 sensor locations since ~2022, each with 4 depths (root zone, 30/60/90 cm), Watermark SWP in kPa; includes a manually derived irrigation overview 2022–ca. 2024 (amounts are estimates, not measurements). Data: provided by LHH, local copy on the project workstation (Rohdaten/Darstellung/Export/Stammdaten + README by LHH, Oct 2025). Per the README the data may only be used for the purpose of the BlueGreenCity-KI project.
• Wesemann, Hack & Backhaus 2025, SSRN preprint (wesemannMultiParameterModelDrought2025) — our multi-parameter drought risk model (48,219 trees, ten weighted criteria); references the LHH vegetation sensors; citable basis for the policy documentation.

Triggers and monitoring:

• Härmann et al. 2025, Hydrological Processes (hormannUrbanTreeDrought2025) — practitioner-focused review of drought stress detection/monitoring; visual inspection detects stress too late.
• Rosenberger et al. 2024, Sustainable Cities and Society (rosenbergerInfluenceAgeSoil2024) — age, soil volume, and climate change drive water availability at urban tree sites; matches the #243-WP8 age classes and the pit_size/tree_disc feature groups.

Amounts and frequency:

• DeGaetano & Hudson 2000, ASCE J. Urban Planning & Dev. (degaetanoSpecificationSoilVolume2000a) and DeGaetano 2000, Arboriculture & Urban Forestry (degaetanoSpecificationSoilVolume2000) — derive irrigation frequency from soil volume and species-specific water demand; the quantitative basis for the amount rules.
• Gobatti et al. 2025, npj Urban Sustainability (gobattiImpactSoilMoisture2025) — maintaining high soil moisture through irrigation produces “no thermal stress” zones; supports cooling targets.
• Nielsen et al. 2007, Arboriculture & Urban Forestry (nielsenSoilWaterDynamics2007) — 100 street-tree pits with different irrigation levels; high pit-to-pit variability of water retention justifies site-specific amounts.

Young trees and species:

• Rahman et al. 2013 (rahmanEffectPitDesign2013) — pit design and soil composition control the establishment period; young-tree rules need pit parameters.
• Smith et al. 2019, PLoS ONE (smithLiveFastYoung2019) — accelerated growth and mortality in street trees; higher care intensity for young classes.
• Hirsch et al. 2023 (hirschDroughtToleranceDiffers2023) — drought tolerance differs strongly between species; no universal scheme.
• Petruzzellis et al. 2022 (petruzzellisTurgorLossPoint2022) — physiological traits predict species-specific decline risk; input design for the species_resilience group.
• Szota et al. 2019 (szotaStreetTreeStormwater2019) — stormwater SCMs reduce runoff but do not benefit established trees; irrigation decisions must not rely on SCMs alone.
• Petrova et al. 2025 (petrovaWhatKillsMature2025) — systematic review: drought is the leading cause of mature street-tree death; the stress-dominated ranking is on track.

Control and planning (methodological relatives of the MILP planner):

• Wei et al. 2026, Landscape and Urban Planning (weiMachineLearningbasedMultiobjective2026) — ML-based multi-objective irrigation optimization (ANN surrogate + genetic algorithm, cooling vs. water-use Pareto front); direct comparison point for #223 evaluation.
• Rebholz et al. 2026, Electronic Markets (rebholzRootingSustainableSmart2026) — co-designed smart-irrigation decision support system for urban green; design pattern for the Taipy review (Phase 3) and LHH collaboration.
• Broadbent et al. 2018, Urban Climate (broadbentCoolingEffectIrrigation2018) — irrigation as a targeted cooling strategy under heatwave conditions; supports heatwave_priority_boost.
• Marchionni et al. 2020, Water Resources Research (marchionniGroundwaterBuffersDrought2020) — groundwater buffers drought effects in urban green; further support for the groundwater_too_high exclusion.

Current Boundary

Resources are currently aggregate daily capacities: water litres, team hours, vehicle count, and maximum jobs per vehicle. Source-specific water quality, depots, refilling, street-network travel times, route sequencing, field approval, and mobile completion reporting are intentionally not fabricated. They require reliable operational inputs and are later extensions around the same planning API.