NASA Space Apps 2026 · Be An Earth System Trend Detective!
An Earth-change intelligence pipeline, from orbit to decision. Three working modules turn NASA satellite data into decisions: detect, confirm, adapt. Each one answers what is changing, where, how much, and whether it's significant.
Three recent events, each seen from orbit, each with a decision that came too late or without the full record.
Fire
14.6 Mhaof Canada's forest burned in the 2023 season, and over 230,000 people were evacuated.
Flood
230confirmed deaths from the October–November 2024 DANA floods in Spain (222 in the Valencia region), across 78 municipalities.
Soil
33%of the world's land is moderately to highly degraded by erosion, compaction, salinization and pollution.
Satellites already see these changes, but the evidence rarely reaches a decision intact. Fire records are split across sensors, so long trends can't be trusted. Optical satellites go blind under the storm clouds that cause floods. And the people choosing what to plant next year rarely get the land's history in a usable form. A single dramatic image can't tell anyone whether something is a trend.
What · 02 · The system
Fire managers can't compare a 2005 fire season with a 2023 one, because NASA's fire record switches sensors in 2012 and VIIRS sees about 5.6 times more fire than MODIS. Without one consistent record, no one can say if fire is really increasing.
8.5% · lower error than the best baseline on 2022–2025 Canada test dataDuring a flood the sky is covered, so optical satellites see cloud, not water. Radar sees through it day and night, but responders and the public can't read radar images, and one flood map can't say if floods are a trend.
3.0× · the water overlap of the standard Otsu threshold on hand-labelled flood chipsA farmer choosing next year's rotation has to weigh soil, water, erosion and a changing climate, but the satellite evidence for their own field is scattered across agencies and formats.
0.8176 · next-season crop accuracy on 66,666 held-out pixels (macro-F1 0.5012)Every module answers the same four questions: what is changing, where, by how much, and is it significant? Mann-Kendall and Sen's slope on 23 to 45 years of NASA records.
Canada fire +3,019 detections/yr (p = 0.035) · Iowa frost-free +0.43 d/yr (p = 0.006)
01 · Detect
What · 02 · The system
01 · Detect
Fire managers can't compare a 2005 fire season with a 2023 one, because NASA's fire record switches sensors in 2012 and VIIRS sees about 5.6 times more fire than MODIS. Without one consistent record, no one can say if fire is really increasing.
8.5% · lower error than the best baseline on 2022–2025 Canada test data
02 · Confirm
During a flood the sky is covered, so optical satellites see cloud, not water. Radar sees through it day and night, but responders and the public can't read radar images, and one flood map can't say if floods are a trend.
3.0× · the water overlap of the standard Otsu threshold on hand-labelled flood chips
03 · Adapt
A farmer choosing next year's rotation has to weigh soil, water, erosion and a changing climate, but the satellite evidence for their own field is scattered across agencies and formats.
0.8176 · next-season crop accuracy on 66,666 held-out pixels (macro-F1 0.5012)
04 · Measure
Every module answers the same four questions: what is changing, where, by how much, and is it significant? Mann-Kendall and Sen's slope on 23 to 45 years of NASA records.
Canada fire +3,019 detections/yr (p = 0.035) · Iowa frost-free +0.43 d/yr (p = 0.006)
What · the modules
Open any module for its live demo, results, figures and data. Every tool runs in your browser from precomputed NASA data.
The problem. Fire managers can't compare a 2005 fire season with a 2023 one, because NASA's fire record switches sensors in 2012 and VIIRS sees about 5.6 times more fire than MODIS. Without one consistent record, no one can say if fire is really increasing.
The problem. During a flood the sky is covered, so optical satellites see cloud, not water. Radar sees through it day and night, but responders and the public can't read radar images, and one flood map can't say if floods are a trend.
The problem. A farmer choosing next year's rotation has to weigh soil, water, erosion and a changing climate, but the satellite evidence for their own field is scattered across agencies and formats.
What · the connective tissue
The Trend Detective layer: pick a variable, get the NASA record, plot it and test whether the change is real.
A steady rise or fall over many years, not a single bad year. We test one value per year.
Compares every pair of years: does the later year tend to be higher (or lower)? It doesn't assume a straight line or a bell curve.
The median change per year across all pairs of years. A single extreme year can't drag it much. We show it with a 95% confidence interval.
Data with no real trend would look this trendy less than 5% of the time. It's evidence of change, not proof of a cause.
Every number comes from our own runs on the public datasets below. Model metrics are one seed each, not a published benchmark.
| Dataset | Module | How it's used | Access |
|---|---|---|---|
| NASA FIRMS MODIS Collection 6.1 (MCD14ML) | Detect | Active fire detections, Terra + Aqua, 2000–2025 | Public, no login |
| NASA FIRMS VIIRS S-NPP 375 m (VNP14IMG) | Detect | Active fire detections, 2012–2025 | Public, no login |
| NASA FIRMS MODIS + VIIRS (trend layer) | Detect | Annual detections and FRP per province, 2003–2025; Mann-Kendall + Sen's slope | Public, no login |
| MODIS MCD64A1.061 burned area (module upgrade) | Detect | Monthly 500 m burned area, tile h11v03, Apr–Oct 2003–2025; Mann-Kendall + Sen's slope | Earthdata login |
| Natural Earth admin-1 boundaries | Detect | Province outlines for the trend map (public domain) | Public, no login |
| NASA-ISRO NISAR L2 GCOV (metadata + browse) | Confirm | Granule search via NASA CMR; public browse images from ASF DAAC | Public, no login |
| Sentinel-1 RTC (ESA Copernicus) | Confirm | VV/VH backscatter, 10 m, Valencia, Sept–Dec 2024; RTC via Microsoft Planetary Computer | Public, no login |
| Sen1Floods11 (CC-BY-4.0) | Confirm | Hand-labelled Sentinel-1 flood chips (Bonafilia et al., 2020) | Public, no login |
| JRC Global Surface Water | Confirm | Permanent-water reference overlay on the test chips | Public, no login |
| NASA POWER daily PRECTOTCORR (trend layer) | Confirm | Daily precipitation 1981–2025 (MERRA-2 based), 35 cells per region; Mann-Kendall + Sen's slope | Public, no login |
| GRACE/GRACE-FO JPL mascon RL06.3 CRI v04 (trend layer) | Confirm | Monthly total water storage 2002–2026, land cells over eastern Spain and southern Ontario; deseasonalised annual means | Earthdata login |
| GPM IMERG V07 monthly GPM_3IMERGM (trend layer) | Confirm | Satellite rainfall, 0.1°, eastern Spain and southern Ontario boxes, 1998–2024: annual, Sep–Nov and wettest-month totals | Earthdata login |
| NASA/JPL OPERA DSWx-S1 (module upgrade) | Confirm | Radar-derived surface water, 30 m, 18 passes 1 Oct–19 Nov 2024 over the Valencia flood zone | Earthdata login |
| NASA POWER | Adapt | Daily climate 2000–2024: temperature, rainfall, root-zone wetness | Public, no login |
| NASADEM | Adapt | Elevation and slope (NASA JPL / LP DAAC) | Public, no login |
| MODIS MOD13Q1 NDVI | Adapt | 16-day vegetation index; bare-soil months | Public, no login |
| USDA NASS Cropland Data Layer | Adapt | 30 m crop maps 2009–2025 (public domain) | Public, no login |
| ISRIC SoilGrids | Adapt | Clay, sand, organic carbon, pH, bulk density (CC-BY-4.0) | Public, no login |
| NASA POWER daily T2M_MIN, T2M_MAX, GWETROOT, PRECTOTCORR (trend layer) | Adapt | 1981–2025 growing-season indicators, central Iowa (35 cells) and Story County; Mann-Kendall + Sen's slope | Public, no login |
| NASA HLS v2.0 HLSL30 + HLSS30 (module upgrade) | Adapt | 30 m NDVI, Story County, two mid-summer scenes (1 Aug 2024, 31 Jul 2025) | Earthdata login |
| NASA SMAP L3 enhanced soil moisture SPL3SMP_E (module upgrade) | Adapt | 9 km surface soil moisture, Story County and central Iowa, May–Sep means 2015–2025; Mann-Kendall + Sen's slope | Earthdata login |
| GLDAS Noah 2.1 monthly GLDAS_NOAH025_M (trend layer) | Adapt | Soil moisture at four depths to 1 m, central Iowa, May–Sep means 2000–2025 | Earthdata login |
Methods
Detect. Empirical-Bayes per-cell VIIRS/MODIS ratio, corrected by LightGBM; conformal intervals at a 90% target; robust z-score flags (|z| > 3.5); MODIS MCD64A1 burned area as a second, area-based record.
Confirm. Sentinel-1 RTC stacks per orbit; U-Net trained on Sen1Floods11 (14 Optuna trials) against Otsu and a global threshold; OPERA DSWx-S1 water counted outside usual water.
Adapt. LightGBM next-season crop model on CDL history plus NASADEM, POWER, NDVI and SoilGrids; FAO-56 water demand; NSGA-II search over four-year rotations; HLS 30 m NDVI and SMAP soil moisture for field-level context.
Measure. One value per year; Mann-Kendall (two-sided, α = 0.05) and Sen's slope with a 95% CI; Hamed-Rao autocorrelation check; a second series wherever one exists (GRACE/GRACE-FO water storage, GPM IMERG rainfall, GLDAS and SMAP soil moisture).
Validation and limits
What holds up
Models are trained on earlier years and tested on later ones (fire 2022–2025, crops 2024–2025) and beat simple baselines. Trends are tested on a second series where one exists.
Where it falls short
Canada's monthly-total fire error is worse than the baseline (39.4% vs 16.7%). Province fire trends hold on the harmonized record but not MODIS alone, so we call them tentative. Trends show change, not cause. Where a second record disagrees, we say so: GPM IMERG satellite rainfall doesn't confirm the reanalysis drop in southern Ontario's rain extremes, so that claim is dropped.
For the judges
One submission to Be An Earth System Trend Detective!, team GeoAI Toronto. Each module page maps its work to the 2026 criteria.
Addresses a real problem, and how many people it could help.
Original idea, or a meaningful improvement on existing ones.
Scientifically grounded, technically feasible and usable.
Answers the challenge directly, with NASA data as a core part.
Clear, engaging story of the challenge, the solution and why it matters.