Climate-based Pre-screening of Self-sustaining Regreening Opportunities in Drylands: A Case Study for Saudi Arabia

This paper presents a scalable, climate-based pre-screening framework using machine learning and remote sensing to identify cost-effective, self-sustaining regreening opportunities in Saudi Arabia's arid drylands, successfully narrowing down national-scale candidates to thirteen priority locations where native vegetation can thrive without intensive irrigation.

The Gist

Imagine trying to turn a barren desert into a lush garden. In many parts of the world, people try to do this by planting trees and watering them heavily. But in places like Saudi Arabia, where water is as rare as gold, this approach often fails. If you water a plant too much, it gets lazy and can't survive when the hose is turned off. Eventually, the garden dies, and the water is wasted.

This paper is like a smart "pre-check" system designed to find the specific spots in the Saudi desert where nature wants to grow back on its own, without needing a constant hose.

Here is how the researchers did it, broken down into simple steps:

1. The Problem: Guessing is Expensive

Usually, to find a good spot to plant trees, you have to drive out there, dig up soil, check the water, and look at the plants. This is slow and expensive. Plus, looking at satellite pictures (which show greenness) can be tricky. In a desert, a small patch of green might just be a farmer's irrigated field, not a natural forest. Or, a patch might look brown but actually have deep roots waiting for rain.

2. The Solution: A "Climate Suitability Score"

The researchers built a digital detective using Machine Learning (a type of computer brain). They taught this computer to look at the weather history of Saudi Arabia and answer one question: "If we planted a native tree here, could it survive on its own?"

• The Training: They showed the computer 230 different "sample spots." Some were lush and green naturally, some were dry deserts, and some were places where humans had ruined the land (like overgrazed areas).
• The Data: Instead of just looking at "is it hot or cold?", the computer analyzed 23 different weather factors (like soil moisture, wind, evaporation, and rainfall) over five years.
• The Result: The computer gave every square inch of Saudi Arabia a Climate Suitability Score (CSS). A high score means the climate is perfect for plants to survive without help. A low score means it's too harsh.

3. The "Sweet Spot" Hunt

Having a high score isn't enough. If a spot is already a lush forest, you don't need to "restore" it. The researchers looked for a specific combination:

• High Climate Score: The weather could support a forest.
• Low Greenness: The land is currently brown or bare.

They called these "Opportunity Zones." These are places where the climate says "Yes, you can grow here," but the land is currently empty, likely because of past damage or overgrazing.

4. Narrowing the List

From a map of the whole country, they found 25 promising spots. But they didn't stop there. They applied a "real-world filter":

• Is it too close to a city? (No, we don't want to fight urban expansion).
• Is it in a volcano field with hard rock? (Maybe not, roots can't grow there).
• Can we actually get a truck there? (Yes, we need access).

After this filter, they were left with 13 priority locations that are ready for field testing.

5. The "Blueprint" for Success

How do they know what the restored land should look like? They used a clever trick: Climatic Analogs.

Imagine you want to build a house in a new town. You look at a house in a nearby town that has the exact same weather and soil. That existing house is your "blueprint."

• The researchers found existing, healthy ecosystems that have the exact same weather as their 13 target spots.
• They measured how green those healthy spots are.
• The Finding: On average, the target spots could support 2.5 times more vegetation than they currently have. This gives them a realistic goal: "We don't need to turn this into a rainforest; we just need to get it to look like that healthy neighbor."

The Bottom Line

This paper doesn't plant the trees yet. Instead, it provides a cost-effective map that tells policymakers and scientists exactly where to look first. By using weather data and computer models, they can skip the expensive guesswork and focus their limited resources on the 13 spots where nature is most likely to say, "Yes, I can grow here on my own."

It's like having a weather forecast that tells you exactly which day to plant your seeds so they don't just wither away.

Technical Summary

Technical Summary: Climate-based Pre-screening of Self-sustaining Regreening Opportunities in Drylands

Problem Statement
Large-scale land restoration in drylands, such as the Saudi Green Initiative, often relies on intensive irrigation to establish vegetation. This approach is unsustainable in water-scarce regions where groundwater replenishment cannot meet agricultural and population demands. A critical challenge is distinguishing locations capable of supporting self-sustaining native vegetation from those that require continuous external inputs. Traditional selection methods relying on the Normalized Difference Vegetation Index (NDVI) are often misleading in arid climates; low NDVI values may indicate degraded land suitable for restoration or simply barren desert, while high NDVI values may reflect irrigated agriculture rather than natural persistence. Furthermore, comprehensive on-site assessments of soil and water quality are costly and difficult to scale. There is a need for a scalable, cost-efficient pre-screening framework that identifies restoration opportunities based solely on climate data to prioritize field campaigns.

Methodology
The authors propose a machine learning-based framework to generate a Climate Suitability Score (CSS) for vegetation survival. The methodology involves the following steps:

1. Data Integration: The study utilizes the ERA5-Land dataset (23 climate variables, 3-hourly resolution from 2020–2024) and multi-year NDVI data (eVIIRS and Sentinel-2) for Saudi Arabia.
2. Training Data Curation: A dataset of 230 expert-curated sample locations was created, categorized into four groups to break the simple correlation between current vegetation and climate suitability:
   • High suitability/High vegetation (persistent natural vegetation).
   • Low suitability/Low vegetation (barren land).
   • Low suitability/High vegetation (irrigated agriculture).
   • High suitability/Low vegetation (degraded land with potential for recovery).
3. Model Architecture: Two complementary machine learning models were trained on compressed climate vectors:
   • Best Linear Unbiased Predictor (BLUP): Uses Fourier coefficient selection for input compression, drawing on methods from quantitative genetics.
   • Neural Network (NN): Uses an autoencoder for latent vector compression followed by a classifier.
   • Ensemble Approach: Both models were evaluated across various input sizes. Since neither model consistently outperformed the other, an ensemble approach (averaging results) was adopted to enhance robustness and avoid overfitting.
4. Opportunity Mapping: The CSS was combined with summer NDVI data to identify "opportunity zones": areas with high climatic suitability but low current vegetation cover. This contrast map filters out irrigated agricultural systems (high NDVI, low CSS) and arid deserts (low NDVI, low CSS).
5. Multi-criteria Filtering: The top 25 candidates were further screened based on elevation, terrain, proximity to settlements, accessibility, and visible vegetation structure using high-resolution imagery, reducing the list to 13 priority locations.
6. Benchmarking: For each candidate, climatically analogous intact ecosystems with higher NDVI were identified to serve as realistic restoration targets and templates for species selection.

Key Results

• Model Performance: The BLUP and Neural Network models showed strong agreement (71% intersection over union) in predicting suitability patterns, validating that the results rely on robust climatic signals rather than algorithmic artifacts.
• Spatial Distribution: The models successfully identified high-suitability regions in the southwestern highlands (Asir, Jazan) and the Red Sea coastal plain, while correctly assigning low suitability to the hyper-arid interior (Rub al-Khali).
• Candidate Identification: The framework narrowed the search space from a national scale to 13 feasible restoration sites spanning diverse climatic zones (Red Sea Coastal, Interior, Northern, Asir Highlands) and landforms (wadis, volcanic substrates, plains).
• Restoration Targets: By comparing candidate locations with climatically analogous native ecosystems, the study estimates that vegetation coverage could realistically increase by an average factor of 2.5 in these areas without intensive irrigation.
• Differentiation: The approach successfully distinguished between naturally persistent vegetation, degraded land with recovery potential, and agriculturally maintained greenery that relies on irrigation.

Significance and Claims
The paper claims to provide a scalable, cost-effective pre-screening tool that directs restoration efforts toward landscapes with higher long-term ecological potential in water-limited regions. By prioritizing sites where climate alone can support vegetation, the framework aims to reduce the risk of project failure and the waste of resources on unsustainable irrigation-dependent projects.

The authors emphasize that their approach:

• Reduces Costs: It minimizes the need for extensive field campaigns by filtering candidates using only global climate data.
• Sets Realistic Expectations: By using intact native ecosystems as benchmarks, it helps define achievable vegetation density targets, preventing over-planting in water-scarce environments.
• Is Transferable: While demonstrated in Saudi Arabia, the framework is applicable to other dryland regions globally (e.g., North Africa, Central Asia, Australia, the US Southwest), provided new training data is collected to account for local microclimates.

The study acknowledges limitations, including the coarse resolution of the ERA5 dataset (9 km) which may miss narrow wadis, the limited size of the training dataset, and the lack of ground-truth soil data. However, the authors argue that the consistency of the model outputs and their alignment with known ecological patterns demonstrate the validity of the CSS approach for guiding initial restoration planning.