The Eden of Apples
This story shows how climate volatility can collapse yield even when the valley still looks healthy, and where adaptation becomes urgent by 2050.
Val di Non, carved into the Italian Alps, produces nearly 1 in 5 apples eaten in Italy. More than 20,000 small parcels, managed by 4,000 farming families in a regional cooperative, make the valley look like a model of agricultural stability.
But that stability depended on a reliable seasonal script. Warmer springs, unstable cold snaps, and more erratic extremes are now breaking the timing that orchards once could trust.
The Secret: The Taste of the Dolomites
The valley's edge came from alpine precision: predictable cold winters, mild summers, and a 15°C daily temperature variation protected by surrounding mountain chains.
That thermal rhythm is what built the valley's famous apple color and crunch. It is also exactly what climate volatility is now destabilizing.
The Satellite's First Clue
In August 2017, the European Space Agency's Sentinel-2 (Copernicus) satellite constellation scanned the valley. It measured the Normalized Difference Vegetation Index (NDVI) — a standardized proxy for photosynthetic health and green biomass. (In plain terms: NDVI is a 0–1 score of how green and photosynthetically active the vegetation is. Near 1.0 = dense healthy canopy. Near 0.2 = bare soil.)
Technical Methodology: Surgical Spectral Recovery
- Source: Multispectral L2A imagery (10m native resolution).
- Frequency: Sentinel revisit is ~5 days, but cloud-screened clean scenes are irregular.
- Precision: Data is Orchard-Masked. We filtered out all noise from forests, rivers, and cliffs using 42,000+ individual apple plot boundaries (GPS).
The satellite confirmed what farmers could see: the trees were vibrantly healthy. In fact, NDVI was actually +2.9% higher than the 10-year historical average.
Decoding the trajectory reveals one confirmed anomaly and one data gap. During Full Bloom, we have no clean satellite scene, so a short bloom shock cannot be directly resolved. The confirmed signal appears later: in Canopy Build, NDVI drops sharply in mid-June, then rebounds into a "Perfect Canopy" by maturation.
Why this matters (click for the evidence)
- Sampling reality: Full Bloom (DOY 110-130) has no usable clean scene in 2017 because we only kept low-cloud acquisitions and excluded cloud-contaminated dates.
- Observed jump: NDVI changes -0.156 from DOY 147 to 164.
- Observed rebound: NDVI changes +0.182 from DOY 164 to 187.
If the valley was greener than ever, the harvest should have been spectacular, right?
But the Trees Were Barren
What the satellite couldn't see was that the vibrant green leaves were hiding a catastrophe. The 2017 harvest was decimated.
Yield crashed by a staggering 63% across the valley. The trees had leaves, but no fruit. They were perfectly healthy, but economically barren.
This is the "Double Anomaly." How does an orchard flourish vegetation while its fruit production collapses entirely?
Cracking the Code with AI
We trained AI to answer one hard question: can it detect a real yield disaster even when the canopy still looks green from space?
Satellite-only signals were not enough. Once we added frost timing, hail stress, and terrain context, the model stopped being fooled by a healthy-looking canopy.
Show model validation details
We used a strict Leave-One-Year-Out setup and fully held out 2017 during training to test whether the model could detect a crisis it had never seen.
| Algorithm | Features | R² Score |
|---|---|---|
| Naive Random Forest | Spectral (NDVI/NDRE) | 0.32 |
| LOGO XGBoost | Spectral + Topography | 0.58 |
| Hybrid AI Fusion | Spectral + Forensic Climate | 0.71 |
Show top model drivers
The highest-impact drivers are mostly timing and stress indicators, not just "how green" the orchard looks:
- Spring green-up timing (27.1%) as proxy for frost exposure.
- Water stress / NDMI (8.9%) for drought pressure inside canopy tissue.
- Hail damage dip (7.1%) for sudden mechanical stress events.
- Heat units / GDD (3.8%) for ripening trajectory.
The Silent Killer: Timing > Temperature
We checked 32 years of ground-truth data from the official MeteoTrentino Station Network and found the key paradox: both 2018 (-6.7°C in March) and 2021 (-6.3°C in early April) were colder than 2017 (-5.0°C in late April), yet they did not trigger the same collapse.
The difference was phenology. In 2017, a false spring pushed blossom opening into mid-April, and then April 18-21 dropped below the -2°C damage threshold and the -4°C catastrophic threshold. Leaves survived, blossoms did not, which is why space still saw green while yield collapsed.
What the radar still saw through the clouds
We saw the 2017 story through Sentinel-2 spectral signals and weather station ground data. The next question is whether Sentinel-1 radar could add a third witness, especially when cloud cover hides the valley during the April frost window.
Click to see what radar could add
- It could reduce the spring sampling gap with parcel-level VV, VH, and change anomalies.
- It could track wetness and surface conditions that shape frost exposure on the valley floor.
- It could compare pre- and post-event structural disturbance after hail when optical scenes are blocked.
Radar is powerful in cloudy periods, but alpine terrain complicates the signal. Slopes, orchard row orientation, layover, and shadow mean Sentinel-1 would work best here as a relative anomaly layer, not as a direct temperature measurement.
The Ghost of Hail
Beyond frost, 2017 was hit by localized "Hail Corridors." These atmospheric bowling alleys can destroy a harvest in minutes, leaving trees lush but fruit shattered.
The Forensic Signature
Unlike frost, hail creates a "Spectral Dip."
Explainer: The Red-Edge Signature
In the chart below, you can see the sharp drop in NDRE (Normalized Difference Red Edge) during the 2017 summer. Unlike NDVI which remains high (lush leaves), NDRE captures the internal chlorophyll stress of the fruit-bearing stems.
The AI model learned to spot these sudden drops. By connecting them across the terrain, we've mapped the recurring "Hail Corridors"—meteorological bowling alleys where landscape topography naturally funnels standard storm cells into icy strikes.
The 2050 Reality & The "Blurry Valley"
The Prediction Challenge
To look beyond 2017, we need future climate scenarios, not just historical evidence. The problem is that climate change in mountain orchards does not arrive as one smooth number. It shows up as earlier springs, unstable cold snaps, and local pockets of stress that can sit inside the same valley on the same night.
This is the "Blurry Valley" problem. A single 5km climate pixel effectively averages the whole basin into one answer. That hides cold-air pooling: on clear nights, dense freezing air drains into the valley floor while nearby slopes stay much warmer. A model predicting a safe 4°C at 5km scale can still miss a lethal -4°C frost pocket inside the orchards below.
Click to unpack CMIP6, projections, and reanalysis
Climate projections are modelled futures under different greenhouse-gas pathways. They are not weather forecasts for one day; they are scenario-based estimates of how conditions shift over decades.
Spatial resolution means the size of each grid cell in the model. A coarse global CMIP6 cell is roughly 100km across, so one cell can blend mountains, valleys, lakes, and plains into one average value.
Downscaling means taking that coarse climate signal and translating it to a finer local grid. In this story, the chain is: global CMIP6 (~100km) -> regional CMCC product (5km) -> our orchard-scale correction (1km).
CMIP6 is the international model framework behind many of those futures. In this project, CMCC provides the regional future baseline that preserves the broad warming trend.
Reanalysis combines observations with physics-based models to reconstruct past atmospheric conditions. Products like ERA-5 help anchor the regional climate field, but even a refined 5km surface still smooths away orchard-scale basin physics.
That is why we still downscale again. Our final step does not invent a new future; it learns the local delta from stations, terrain, and cold-air pooling, then applies that correction inside each coarse regional pixel.
Spin the Frost Pocket
Drag to rotate this basin. The mesh is now a real Copernicus GLO-30 terrain cutout centered on the strongest historical frost pocket in Val di Non, with real orchard plots dropped onto the slope.
Loading the 3D comparison...
The AI Downscaling Solution
We keep the future warming trend from CMCC. The AI does one smaller, more useful job: it learns how each part of Val di Non behaves differently inside that coarse 5km block.
For the casual reader, the story is simple: stations teach the local correction, history brings back spring volatility, and terrain shows where cold air pools. If you want the method, click the steps below.
Click a step to reveal one layer of the 1km map. The technical detail stays hidden unless you open it.
Anchor the valley to local truth
Loading the interactive Act 4 explainer...
More detail
That is why the final surface is not just “sharper.” It is more physical. The 5km warming trend stays intact, but each 1km orchard cell gets its own microclimate behavior, which is why frost pockets and water stress can sit inside the same regional pixel.
The Threat Matrix
Quick read: every 1km cell gets one risk score by combining four threats - frost, drought, instability, and hail.
The red and orange cells are where these threats overlap. That is where intervention pays back fastest.
Start with the orchards or communes you actually manage.
See whether frost, drought, hail, or volatility is pushing the score.
Use the highest-overlap cells as the first budget and adaptation targets.
Show scoring method
The baseline uses the 2016-2024 window. For each cell, the score blends four components with equal weights (25% each):
- Frost Frequency: how often bloom timing overlaps with lethal freeze windows.
- Chronic Water Stress: persistent NDWI decline indicating long-term drying pressure.
- Yield Volatility: unstable production behavior across the multi-year cycle.
- Hail Intensity: recurrence of spectral-detected mechanical strike corridors.
Strategic Implications
Use the planner to compare one red cell and one orange cell, then toggle netting, irrigation, and frost fans to see which lever changes risk fastest.
The goal is not to erase all risk. It is to spend first where protection, water management, or upslope transition changes the outcome most.
Use netting where repeat strike paths already appear in the record.
Use irrigation where drought and year-to-year instability stack together.
Plan varietal and elevation change where spring losses keep repeating.
Show planning rationale
For the most volatile low-altitude zones, growers and planners are already evaluating a shift upslope by roughly 300 m to preserve the diurnal thermal profile that supports fruit quality.
The era selector and adaptation planner are unlocked in the next section — scroll down to compare 2040, 2070, and 2100 risk scenarios and test mitigation strategies live on the map.
Map & Risk Scenario Explorer
Use the Era Selector (bottom-left of map) to project how the risk landscape shifts by 2040, 2070, and 2100. The map recolours in real time based on downscaled CMCC pathways.
Use the Adaptation Planner (top-right of map) to test whether hail netting, precision irrigation, or frost fan installation changes the risk score for your cells.
See which cells transition from orange to red under the worst-case pathway.
Tick hail netting in the planner and watch red hail-corridor cells cool down.
The hover panel shows the four component scores so you can identify the dominant driver.
Methodology & References
This project was built following the Martini Glass structural methodology (Segel & Heer, 2010), transitioning from an author-driven forensic narrative to a reader-driven strategic exploration tool.
Academic References
- Kosara, R., & Jock, M. (2013). Storytelling: The Next Step for Visualization. IEEE Computer.
- Segel, E., & Heer, J. (2010). Narrative Visualization: Telling Stories with Data. IEEE TVCG.
- Donoho, D. (2017). 50 Years of Data Science. Journal of Comp. and Graph. Stats.
Data Provenance
- Satellite: Sentinel-2 multispectral (ESA/Copernicus).
- Topography: Copernicus GLO-30 DEM basin cutout for the Act 4 inversion inset.
- Future Models: CMCC CMIP6 EC-Earth3-Veg (SSP3-7.0 scenario).
- Agronomics: Rigenera Orchard Metadata & Historical Yield Registries.
- Climate: ERA-5 Reanalysis & MeteoTrentino Local Station Network.