Reciprocal Environmental Decision Support (REDS): better tailored advice in return for data

This paper introduces Reciprocal Environmental Decision Support (REDS), a novel system that synergizes data flows by using user-contributed citizen science data to simultaneously provide tailored environmental advice and automatically refine the underlying predictive models, as demonstrated by a proof-of-concept Garden Advice tool that successfully improved House Sparrow presence predictions through automated Bayesian updating.

The Gist

The Big Idea: A Two-Way Street for Nature Advice

Imagine you have a GPS app for your car.

• Old Way (One-Way): You tell the GPS where you are, and it tells you the fastest route. But the GPS doesn't learn from your drive. If you hit a traffic jam, the GPS doesn't know about it until the next person tells them.
• The New Way (REDS): You tell the GPS where you are, and it gives you a route. But as you drive, your speed and traffic data automatically update the map for everyone else. The next person gets a better route because of your drive, and you get even better advice next time because the map got smarter.

This paper introduces a system called REDS (Reciprocal Environmental Decision Support). It's a "two-way street" for nature advice. Instead of just giving you advice, the system learns from your answers to give better advice to you and everyone else later.

---

The Problem: The Broken Loop

Currently, there are two types of nature apps:

1. The "Expert" App (EDS): You ask, "How do I make my garden better for birds?" The app gives you a generic answer based on old data. It doesn't learn from your specific garden.
2. The "Citizen" App (Citizen Science): You report, "I saw a bird here!" The scientists use your data to update their big models. But you don't get any personalized advice back. You just give data and leave.

The Gap: The "Expert" app doesn't get smarter from your garden, and the "Citizen" app doesn't help you make decisions.

The Solution: The "Garden Advice" App

The researchers built a prototype app called Garden Advice to test if they could close this loop. They focused on the House Sparrow, a common bird that is actually in decline.

How it worked (The Analogy):
Imagine the app is a smart tutor teaching you about sparrows.

1. The Starting Point: The tutor started with a textbook written by professional scientists (data from Glasgow). It knew the basics but wasn't perfect for every garden in the UK.
2. The Lesson: You (the user) opened the app and drew a map of your garden. You told the app: "I have a big lawn, a hedge, and I see sparrows near the roof."
3. The Exchange:
   • You get advice: The app said, "Based on your map, you have a 70% chance of seeing sparrows. If you cut your grass shorter, that chance goes up!"
   • The app learns: The app took your data and instantly updated its "textbook." It realized, "Oh, in gardens like this, grass is actually good for sparrows, not bad!"
4. The Result: The very next person who used the app got advice based on your garden data, making the advice more accurate for everyone.

What Did They Find?

1. The "Smart Tutor" got smarter.
When they tested the app, the version that had learned from real users predicted bird sightings better than the version that only used the old textbook. Even though the users weren't professional bird watchers, their collective data was good enough to improve the model.

2. It corrected a "Bias."
The old textbook thought sparrows loved being close to roofs. But when regular people started using the app, the model changed its mind. Why? Because people looking at roofs from inside their houses can't see the birds well! The new data showed that sparrows actually like grass and hedges more than the old model thought. The system corrected the "blind spot" of the experts.

3. People actually used it.
Even though people were paid a small amount to try the app, many of them stayed to see the results. Over half of them checked their "Sparrow Score" and even used the "planning mode" to see how changing their garden (like adding a hedge) would help the birds.

Why Does This Matter?

Think of this like a community kitchen.

• In the past, you either cooked a meal for yourself (Decision Support) or you gave your ingredients to a chef to make a soup for everyone (Citizen Science).
• REDS is like a potluck where everyone brings a dish, but the recipe book updates in real-time. If you bring a great potato salad, the recipe book changes so the next person knows exactly how to make a potato salad.

The Takeaway:
This study proves that we don't have to choose between "getting advice" and "giving data." We can do both at the same time. By creating a system where your contribution immediately improves the advice you get, we can build a smarter, more accurate understanding of nature that helps us all make better decisions for the environment.

In short: You help the system, and the system helps you back. It's a win-win for nature.

Technical Summary

1. Problem Statement

Environmental Decision Support (EDS) systems and Citizen Science (CS) systems currently operate with unidirectional data flows:

• EDS: Data flows from the user to the system to generate tailored predictions (System $\to$ User). However, the rich input data provided by users is typically discarded after the query is answered and not used to improve the underlying models.
• CS: Data flows from the user to the system to improve global models (User $\to$ System). However, these systems rarely provide immediate, tailored feedback or decision support to the individual contributor based on their specific local data.

The Gap: There is a lack of systems that close the information loop by simultaneously providing tailored advice to users and using their contributions to iteratively update the predictive models in real-time. Existing systems like iNaturalist (model updates but no EDS) or Cool Farm Tool (EDS but no model updates) fail to achieve this synergy. Furthermore, existing EDS models are often static and may not reflect current local realities or changing environmental conditions.

2. Methodology

System Design: Garden Advice (GA)
The authors developed "Garden Advice," a proof-of-concept REDS system designed for UK domestic gardens.

• Target Species: House Sparrow (Passer domesticus), chosen as an indicator species due to its global decline and association with human habitats.
• Data Collection: 71 UK residents were recruited via Prolific to map their gardens using the GA app. Participants drew habitat boundaries (trees, bushes, grass, roofs, etc.) and recorded specific areas where they observed sparrows.
• Model Architecture:
  • Base Model: A Bayesian binary logistic regression model initially parameterized using the Glasgow House Sparrow Project (GHSP) dataset (trained observer data).
  • Predictors: Proportions of six habitat types (tree, bush, hedge, grass, roof, artificial surface) and two distance variables (proximity to hedge/roof).
  • Structure: The model includes fixed effects for predictors and random effects (intercepts and slopes) for location to account for spatial autocorrelation and inter-location heterogeneity.
  • Updating Mechanism: The system employs automated Bayesian updating. As new user data is contributed, the model parameters are updated daily. The system uses a prevalence-based offset to reconcile the case-control design of the GHSP data with the full presence-absence data from GA.
  • Inference: While preregistered for Markov chain Monte Carlo (MCMC), the study utilized Integrated Nested Laplace Approximation (INLA) for faster cross-validation due to computational constraints.

Evaluation Strategy

• Primary Hypothesis: The model updated with user-contributed GA data would predict new GA data better than the initial GHSP-only model.
• Metrics: Prediction accuracy was evaluated using Brier scores (Mean Squared Error), decomposed into Discrimination (resolution) and Miscalibration (calibration).
• Validation: Leave-one-group-out cross-validation (hierarchical LOO-CV) was used to ensure models were not evaluated on data they were trained with.

3. Key Contributions

1. Definition and Implementation of REDS: The paper formally defines Reciprocal Environmental Decision Support (REDS) as a system type that closes the information loop, allowing user observations to update models that generate advice for those same users. It identifies Garden Advice as the first REDS system specifically for ecosystem management based on environmental conditions.
2. Demonstration of Untrained Data Utility: The study proves that data from untrained, non-expert observers (citizens) can be sufficiently high-quality to refine models initially built on expert data.
3. Dynamic Model Updating: The system successfully demonstrated automated, daily Bayesian updating, showing that models can evolve rapidly (within days) to reflect new data, unlike traditional static EDS models.
4. Effect Consistency Metric: The authors introduced a novel way to present model results for decision support: Effect Consistency. This metric estimates the probability that a specific habitat effect (e.g., trees increasing sparrow presence) applies to a random location, separating the magnitude of the effect from its heterogeneity across locations.

4. Results

Model Performance

• Prediction Accuracy: The updated model (GHSP + GA data) significantly outperformed the initial GHSP-only model when predicting new GA data. The Brier score improved by approximately 0.01, corresponding to a non-trivial improvement in probability predictions (~0.1 average improvement).
• Decomposition: The improvement was primarily driven by better discrimination (the model's ability to separate presence from absence), rather than calibration.
• Generalizability: Interestingly, the updated model's ability to predict the original GHSP data decreased slightly, suggesting the new model is better tailored to the broader UK population represented by the citizen data rather than the specific Glasgow cohort.

Parameter Updates and Ecological Insights

• Grass Association: The model shifted from a weakly confident negative association with grass to a positive association (91% probability of real difference). This aligns with literature suggesting sparrows forage on seeds/invertebrates in grass, resolving conflicting previous findings.
• Roof Proximity: The initial strong positive association with roof proximity (from GHSP data) was significantly reduced in the updated model (98% probability of real difference). The authors attribute this to observation bias: sparrows on roofs are harder for untrained observers to see from indoors, leading to under-reporting in the initial expert data or over-reporting in specific contexts.
• Vegetation: The model became more confident in the positive effects of trees, bushes, and hedges.

User Behavior

• Despite being paid to map gardens, 76% of participants voluntarily examined their sparrow or carbon scores.
• 18% used the "planning mode" to simulate garden changes, and 11% examined how those changes affected scores, demonstrating the system's utility as a decision support tool even without explicit instruction to use it.

5. Significance and Implications

• Synergy of CS and EDS: The study validates that combining Citizen Science and EDS creates a "virtuous circle": users provide data to improve the model, and the improved model provides better, more tailored advice to future users.
• Adaptive Management: REDS systems can remain up-to-date with changing environmental conditions (e.g., climate, seasonal shifts) by continuously integrating new data, overcoming the stagnation of static models.
• Scalability and Cost: The automated nature of REDS reduces the burden on scientists to manually analyze and report data, potentially saving time and resources while increasing stakeholder engagement through immediate feedback.
• Future Directions: The authors suggest that REDS could evolve into "digital twins" for ecosystem management. Future work must focus on independent verification of citizen data quality and expanding the approach to more complex demographic models (tracking species health, not just presence).

In conclusion, the paper establishes REDS as a viable and powerful framework for environmental management, demonstrating that reciprocal data flows can yield more accurate, localized, and dynamic decision support than traditional unidirectional systems.