EcoDri System: Drought & Environmental Tech

• EcoDri System is a comprehensive framework combining modular data pipelines, integrated drought indices, and machine learning to forecast environmental impacts.
• It uses advanced spatial-temporal feature extraction with neighbor aggregation and imbalanced class handling to achieve high-accuracy drought predictions.
• Its versatile applications, including solar-driven textile drying and humidity-swing CO₂ capture, underline its role in sustainable, field-adaptable environmental monitoring.

EcoDri System broadly refers to a suite of systems and methodological pipelines designed for data-driven environmental monitoring and decision-support, with concrete instances in drought impact forecasting, solar-driven textile drying, and direct air capture (DAC) of CO₂. The most rigorously characterized and widely cited implementation is the Ecological Drought Information Communication System (EcoDri), which provides operational, spatially explicit, short-term drought impact forecasts using supervised machine learning models and integrated environmental datasets (Geli et al., 20 Dec 2025). Additional references to "EcoDri" or EcoDri-style systems occur in the solar evaporation and DAC literature, describing highly efficient textile-based evaporators (Peng et al., 2018) and humidity-swing CO₂ capture platforms (Shi et al., 2017). This entry focuses primarily on the architecture, methods, and performance of the drought impact forecasting system, but references physically related EcoDri-style implementations for completeness.

1. System Architecture and Data Integration

EcoDri is organized around a modular, four-stage data pipeline for ecological drought impact prediction (Geli et al., 20 Dec 2025):

1. Data Ingestion Inputs are weekly, county-level time series from three domains:
   • Drought Severity and Coverage Index (DSCI): US Drought Monitor, mapped to administrative units.
   • Evaporative Stress Index (ESI): MODIS-ALEXI fPET anomalies, converted to standardized scores per county.
   • Drought Impact Reporter (DIR): Curated text-mined weekly drought impact records categorized by sector (Agriculture, Water, Fire, Plants & Wildlife, Relief, Society, Tourism).
2. Preprocessing and Augmentation
   • Spatial alignment: raster-to-county mapping; aggregation of pixel-level ESI and DSCI to county means.
   • Temporal alignment: resampling and interpolation to weekly granularity.
   • Impact encoding: For each (county, week, impact-category), binary target label $Y = 1$ if at least one DIR record is present.
   • Neighbor-based horizontal aggregation: features assembled for adjacent counties to capture spatial dependencies.
   • Data balancing: Borderline-SMOTE oversampling of impact=1 minority, followed by ENN undersampling of majority class.
3. Model Development
   • For each county, week, impact category, and $W \in \{1,\dots,8\}$ (number of preceding weeks), a feature set is constructed: $$X = \{\mathrm{DSCI}, \mathrm{ESI}, \mathrm{ImpactBinary}\}_{\mathrm{target},\ \mathrm{neighbors}}^{t-W\ldots t-1}$$.
   • XGBoost (XGBClassifier, binary:logistic objective, $n_\mathrm{estimators}=100$, $\mathrm{max\_depth}=6$, $\mathrm{learning\_rate}=0.3$) is trained using an 80/20 temporal split, or leave-one-county-out cross-validation.
4. Forecast Generation and Decision Support
   • Weekly, for each county, features from the previous $W$ weeks for both the target and neighbors are input to the trained XGBoost model to forecast $P(\mathrm{impact}=1)$ for up to 8 weeks ahead.
   • Outputs are mapped to dashboards, threshold-triggered alerts (e.g., for fire agencies), and PDF decision reports.

A schematic pseudocode of the Drought Impacts Prediction (DIP) Framework is reproduced in the reference (Geli et al., 20 Dec 2025).

2. Drought Indices and Impact Taxonomy

2.1 Drought Severity and Coverage Index (DSCI)

DSCI transforms the categorical US Drought Monitor map into a continuous index: $$\mathrm{DSCI} = \sum_{k=0}^4 (k+1) \frac{A_k}{100} \times 100$$ where $A_k$ is the percent area under drought category $D_k$ $(k=0\ldots4)$. DSCI ranges 0–500.

2.2 Evaporative Stress Index (ESI)

ESI is the standardized anomaly of the instantaneous ratio $fPET = \frac{ET}{PET}$ (actual/potential evapotranspiration) per: $$\mathrm{ESI}_t = \frac{fPET_t - \overline{fPET}}{\sigma_{fPET}}$$ ESI captures physical soil-moisture anomalies, operationalized at $\sim4$ km/pixel, with negative departures indicating drought.

2.3 Drought Impact Reporter (DIR) Categories

Category	Definition
Agriculture	Crop and livestock productivity losses, forest impacts
Water Supply & Quality	Reservoir levels, restrictions, water quality impairments
Fire	Wildfires, fire danger restrictions
Plants & Wildlife	Native vegetation/wildlife mortality, habitat disruption
Relief & Response	Disaster declarations, aid, water bans
Society & Public Health	Health advisories, economic impacts
Tourism & Recreation	Recreation sector closures, visitor declines

Impacts are encoded as binary labels per county-week-category.

3. Machine Learning Methodology

EcoDri’s model development is characterized by:

• Temporal–Spatial Feature Expansion: Features include not only the target county’s DSCI, ESI, and recent impacts over $W$ past weeks, but also the corresponding features from spatially adjacent counties.
• Imbalanced Class Handling: Severe class imbalance (impacts are rare events) is addressed using Borderline-SMOTE for synthesizing minority samples and ENN (Edited Nearest Neighbors) for majority undersampling.
• Classifier and Hyperparameters:

XGBoost (XGBClassifier, binary:logistic) is selected for its ability to handle missing values (due to variable neighbor set sizes), default robust hyperparameters, and superior performance over both RandomForest and LSTM baselines.

Training and Validation Strategy

• Spatial Cross-Validation: Leave-one-county-out; at each fold, a single county is held out as the test set, with the remaining counties used for training, augmented on training only.
• Temporal Holdout: Within each fold, the last 20% of weeks in time form the validation set to prevent temporal leakage.

Predictive Targets

• The system forecasts, for each impact category, the probability $P(\mathrm{impact}=1)$ per county and per week, for lead times up to 8 weeks.

Performance Metrics

• Primary metric: F1-score (harmonic mean of precision and recall for binary event/no-event class).
• Models with $\mathrm{F1}\geq0.70$ are deemed good, $\geq0.50$ acceptable (Geli et al., 20 Dec 2025).

4. Operational Performance and Use Cases

4.1 Predictive Skill

At the State (New Mexico) level:

• XGBoost (DSCI+ESI) yields $F1=0.85$–$0.96$ for Agriculture, Water, Fire, and Relief categories up to 8 weeks ahead.
• For Plants and Society impacts, F1 drops from $0.83 \to 0.60$ and $0.82 \to 0.59$, respectively, with increasing lead time.
• ESI alone outperforms DSCI for rapid-onset events, while the combination improves prediction for physical impacts.

At the County level (e.g., Bernalillo County):

• F1 = 0.93–0.99 for major categories (Agriculture, Fire, Relief) up to 4 weeks ahead.
• XGBoost outperforms LSTM and RandomForest, especially given sparse, noisy impact data.

4.2 Decision-Support Integration

• Early Warning: Weekly forecast maps and charts are served via dashboards (EcoDri Portal), supporting localized, actionable drought risk monitoring.
• Alerting: Automated triggers (e.g., $P(\mathrm{Fire~impact}) >0.8$) prompt direct notifications to relevant agencies.
• Stakeholder Feedback: Verification and ground-truthing are facilitated via the DIR input pipeline, enabling continuous model adjustment.

This design allows transition from monitoring drought conditions to forecasting probabilistic impacts, enhancing adaptive capacity of land managers and agencies (Geli et al., 20 Dec 2025).

5. Physical EcoDri-Style Implementations

The EcoDri name is also associated with independently developed, physically engineered systems for environmental mitigation:

5.1 Solar-Driven Textile Dryer

• Architecture: Flexible carbon-black-coated linen wick (central panel with capillary-fed “wings”) suspended above a water reservoir within a foldable, thermally insulated frame (Peng et al., 2018).
• Performance:
  • Evaporation flux $$\dot{m} \approx 1.3~\mathrm{kg}\cdot\mathrm{m}^{-2}\cdot\mathrm{h}^{-1}$$ under 1 sun (1 kW/m²), efficiency $\eta \approx 87\%$.
  • Rapid drying of textiles (e.g., 500 g towel with 1.5 kg water dries in <1.5 h per 1 m² module).
  • Modularity and portability for off-grid deployment.

5.2 Humidity-Swing CO₂ Capture

• Mechanism: Nanoporous materials loaded with carbonate salts exhibit reversible CO₂ uptake based on ambient humidity:
  • Dry state (RH ≲ 20%): Generation of OH⁻, rapid CO₂ absorption, capacity ~0.7 mmol·g⁻¹.
  • Wet state (RH ≳ 60%): Back-reaction releases CO₂ at ambient partial pressure via hydration-triggered free energy swing (Shi et al., 2017).
• Design: Hydrophobic nanoporous scaffolds (e.g., graphene aerogels), cation functionalization at atomic spacing, optimized for thermodynamic favorability of hydrolysis at chosen hydration levels.

The unifying theme is modular, low-cost, and field-adaptable engineering for sustainable evaporation or gas processing, with analytical models for energy and mass balance provided (Peng et al., 2018, Shi et al., 2017).

6. Limitations and Prospective Extensions

6.1 Drought Forecasting System

Limitations:

• Dependence on DIR records introduces spatial/reporting bias; underreporting is significant in low-population areas.
• DSCI’s reliance on human expert input can lag flash events.
• Binary impact quantification cannot isolate severity within a given week; “soft” impacts (e.g., societal) remain challenging (F1 < 0.7).

Future Directions:

• Incorporation of additional indices (SPEI, soil moisture from SMAP/SPI) to enrich features.
• Deployment of attention-based temporal models (e.g., transformer architectures) to capture long-range dependencies.
• Geographic extension to non-US regions; retraining for international DIR analogues.
• Augmentation with socio-economic exposure layers for impact quantification beyond binary occurrence.
• Automated incremental model retraining as new data arrive (Geli et al., 20 Dec 2025).

6.2 Physical Systems

Constraints:

• Textile modules’ performance declines at high ambient humidity (efficiency decreases by ~10–20% for RH > 80%).
• CO₂ DAC systems’ kinetics limited by mass transfer in realistic conditions; full scalability requires robust anti-leaching supports (Shi et al., 2017).

7. Comparative Summary Table

EcoDri System	Domain	Core Methodology	Peak Metric
Drought Forecasting	Environmental Informatics	DSCI/ESI + XGBoost ML	F1=0.85–0.96 (key categories)
Textile Dryer	Solar-driven Evaporation	Carbon-black linen wick, AES	$\eta$=87%, $\dot{m}$=1.3 kg·m⁻²·h⁻¹
CO₂ Capture	Direct Air Capture	Humidity-swing, nanoporous	0.7 mmol·g⁻¹ per cycle

The term EcoDri thus encompasses a modular informatics system for ecological drought forecasting (Geli et al., 20 Dec 2025), blueprint designs for portable solar evaporators (Peng et al., 2018), and engineered platforms for humidity-driven DAC (Shi et al., 2017), all unified by their focus on efficient, data-driven, and field-adaptable environmental mitigation.