Trend per decade (Theil-Sen)
—
Memory and significance
— Hamed-Rao p
Largest shift (Pettitt)
—
Anomalous months and ENSO
— months (≥2 of 3 detectors)
Monthly anomalies (value minus that month's normal) with two trend lines and the largest level shift
Monthly anomaly
Theil-Sen trend
OLS trend
Pettitt step
Anomalous month
Annual means
Averaging each year removes most month-to-month memory
Normal year (mean ± 1σ per month)
The seasonal cycle removed before looking for trends
All six variables at once
Testing six variables raises the chance of a false alarm, so q-values control the false discovery rate. Click a row to open that variable.
| Variable | Trend / decade | OLS / decade | Lag-1 r₁ | n_eff | Naive p | Hamed-Rao p | FDR q | Annual p | Pettitt shift | ONI r | Is it real? |
|---|
Trend by season
Which season drives the change (December counts with the following January–February)
| Season | Trend / decade | OLS / decade | Hamed-Rao p | Signal |
|---|
Months in two dimensions (PCA) and regimes
What explains root-zone soil moisture?
Random Forest on this month's and earlier rainfall, temperature, humidity, radiation and ENSO
Months that at least two of three detectors call unusual
Isolation Forest, Local Outlier Factor and Mahalanobis distance, each with its own data-driven threshold
| Month | Ranking score | Detectors | Mahalanobis D | ONI | What was unusual |
|---|
Trend with and without the anomalous months
Per decade, full period of each variable
Outlier test (Acker et al., 2009)
| Variable | All months | p (all) | Without anomalies | p (without) | Change | Diagnosis |
|---|
Compare two regions
Standardized anomalies of both regions next to the ENSO index. Correlations use an effective sample size that accounts for month-to-month memory.
Trend map over Ecuador (NASA POWER grid, ~55 km cells)
Solid cells stay significant after controlling the false discovery rate across all cells. Pins are the benchmark regions.
16-day NDVI from Terra/Aqua MODIS (MOD13Q1, 250 m), a single pixel at the region's point
Check any coordinate
Downloads NASA POWER for a point and runs the same trend tests
Enter coordinates or turn on probe mode and click the map.
1 km model snapshots (Meteomatics, 2023–2024)
One day every 10 days from a commercial downscaled model; not observations and not monthly means
Data provenance: MODIS granules
Granules of the modis-13Q1-061 collection that contain the point (Microsoft Planetary Computer STAC)
| Granule | Date | Tile | Platform |
|---|
These benchmarks come from train_models.py, run offline. Every model is compared with persistence: predicting that next month repeats this month's anomaly. A model is only useful where it beats that baseline.
Nowcast over the Ecuador grid
Uses same-month weather, so it explains the present rather than predicting the future
| Model | CV R² (dev.) | Test R² | Gain vs. persistence | Test RMSE (σ) | CV folds |
|---|
What the best nowcast relies on
Drop in test R² when each input is shuffled
| Main input | Drop in test R² |
|---|
Forecast: only data known in advance
Predicting soil moisture 1 and 3 months ahead
| Horizon | Model | CV R² (dev.) | Test R² | Gain vs. persistence |
|---|
Regional driver models
The model behind "What explains root-zone soil moisture?", per region. Negative R² means worse than predicting the average.
| Region | CV R² (dev.) | Test R² | Persistence test R² | Regimes (k) | Silhouette | Main input |
|---|
How to use TerraLens
1Ask a question
Pick a region, a variable and a period in the top bar. The card at the top answers in one sentence whether there is a real long-term trend and why.
Or open a guided case in the sidebar: it sets everything up and walks you through the evidence step by step.
2The two filters
- 2001 sensor filter. NASA POWER's solar irradiance comes from GEWEX SRB until 2000 and from CERES SYN1deg from January 2001. Joining both can create a false trend. With the filter on, solar irradiance is analysed from 2001 only.
- Remove extreme months. Recomputes the trend without the months that at least two of the three anomaly detectors flag. If the trend survives, a few extreme events do not explain it (the outlier test of Acker et al., 2009).
3Why the corrections matter
- Memory (autocorrelation). Soil moisture this month looks like last month. Treating 540 months as independent makes weak trends look significant. Hamed-Rao and Santer corrections fix the p-value; the annual-mean test is a second check.
- Steps vs. trends. Pettitt finds the largest level shift, and its p-value comes from a block bootstrap that respects memory. A model comparison then says whether one step or a gradual trend fits better. A trend that is really one step is flagged.
- Many tests at once. Testing six variables, or 96 map cells, produces false alarms by chance. q-values (Benjamini-Hochberg) control that.
4Expert mode and sharing
Expert mode shows every statistical column (naive p, n_eff, annual p, Pettitt dates, ONI correlations, detector details).
The page address keeps your region, variable, period, filters, view and case, so you can share exactly what you are looking at. CSV downloads the monthly series with its metadata.
1Haz una pregunta
Elige una región, una variable y un periodo en la barra superior. La tarjeta de arriba responde en una frase si hay una tendencia real de largo plazo y por qué.
O abre un caso guiado en la barra lateral: prepara todo y te lleva paso a paso por la evidencia.
2Los dos filtros
- Filtro de sensor 2001. La irradiancia solar de NASA POWER viene de GEWEX SRB hasta 2000 y de CERES SYN1deg desde enero de 2001. Unir ambas fuentes puede crear una tendencia falsa. Con el filtro activo, la irradiancia solar se analiza solo desde 2001.
- Quitar meses extremos. Recalcula la tendencia sin los meses que al menos dos de los tres detectores marcan como anómalos. Si la tendencia sobrevive, no la explican unos pocos eventos extremos (la prueba de extremos de Acker et al., 2009).
3Por qué importan las correcciones
- Memoria (autocorrelación). La humedad del suelo de este mes se parece a la del anterior. Tratar 540 meses como independientes hace que tendencias débiles parezcan significativas. Las correcciones de Hamed-Rao y Santer arreglan el p-valor; la prueba de promedios anuales es un segundo control.
- Saltos vs. tendencias. Pettitt encuentra el mayor cambio de nivel y su p-valor sale de un bootstrap por bloques que respeta la memoria. Luego una comparación de modelos dice si se ajusta mejor un salto o una tendencia gradual. Una tendencia que en realidad es un salto queda marcada.
- Muchas pruebas a la vez. Probar seis variables, o 96 celdas del mapa, produce falsas alarmas por azar. Los q-valores (Benjamini-Hochberg) lo controlan.
4Modo experto y compartir
El modo experto muestra todas las columnas estadísticas (p ingenuo, n_eff, p anual, fechas de Pettitt, correlaciones con ONI, detalle de los detectores).
La dirección de la página guarda región, variable, periodo, filtros, vista y caso, así puedes compartir exactamente lo que estás viendo. CSV descarga la serie mensual con sus metadatos.
About TerraLens and its limits
NASA Space Apps 2026 · Proyecto Hail Mary1The problem
A straight line fitted to a satellite or reanalysis series can report a trend that is not there. Three traps cause most false trends:
- Memory: persistent variables such as soil moisture inflate significance.
- Data changes: a new sensor or new input data creates a step that looks like climate change.
- Extremes: one huge El Niño or drought near the end of the record can tilt the line.
2What TerraLens does
- Monthly anomalies, annual means and seasonal subsets, following Acker et al. (2009).
- Theil-Sen slopes with Hamed-Rao and Santer corrections, Pettitt change points with a block bootstrap, a step-versus-trend check and false discovery rate control.
- Three anomaly detectors, an outlier sensitivity test, ENSO correlations with an effective sample size, and a live probe for any coordinate.
- Every number in the guided cases is computed by the app when it starts.
3Data
- NASA POWER monthly points and a grid over Ecuador, 1981–2025: MERRA-2 temperature, humidity, soil wetness and bias-corrected precipitation; GEWEX SRB and CERES SYN1deg solar irradiance.
- NOAA CPC Oceanic Niño Index.
- MODIS MOD13Q1 NDVI (ORNL DAAC), Meteomatics 1 km snapshots and Planetary Computer STAC metadata.
4Limitations
- NASA POWER cells are ~55 km (solar ~110 km). One value mixes very different terrain, especially in the Andes.
- MERRA-2 is a reanalysis, not station data; changes in its inputs can create steps (see the Napo case).
- Each region is a single point. The Guayas and Manaus points fall on cities and rivers, so their MODIS NDVI describes water and urban land, not crops or forest.
- Meteomatics values are model snapshots every 10 days for 2023–2024 only.
- The soil-moisture models barely beat persistence in several regions and fail in the Arctic (negative test R²); the cause has not been isolated.
1El problema
Una recta ajustada a una serie satelital o de reanálisis puede reportar una tendencia que no existe. Tres trampas causan la mayoría de las tendencias falsas:
- Memoria: variables persistentes como la humedad del suelo inflan la significancia.
- Cambios en los datos: un sensor nuevo o nuevos datos de entrada crean un salto que parece cambio climático.
- Extremos: un El Niño o una sequía enormes cerca del final del registro pueden inclinar la recta.
2Qué hace TerraLens
- Anomalías mensuales, promedios anuales y subseries estacionales, siguiendo a Acker et al. (2009).
- Pendientes Theil-Sen con correcciones de Hamed-Rao y Santer, quiebres de Pettitt con bootstrap por bloques, una comparación salto-vs-tendencia y control de falsos descubrimientos.
- Tres detectores de anomalías, una prueba de sensibilidad a extremos, correlaciones con ENSO con tamaño efectivo de muestra y una sonda en vivo para cualquier coordenada.
- Todos los números de los casos guiados los calcula la app al arrancar.
3Datos
- NASA POWER: puntos mensuales y una malla sobre Ecuador, 1981–2025: temperatura, humedad, humedad del suelo y precipitación corregida de MERRA-2; irradiancia solar de GEWEX SRB y CERES SYN1deg.
- Índice Oceánico de El Niño de NOAA CPC.
- NDVI MODIS MOD13Q1 (ORNL DAAC), fotos de Meteomatics a 1 km y metadatos STAC de Planetary Computer.
4Limitaciones
- Las celdas de NASA POWER miden ~55 km (la solar ~110 km). Un valor mezcla relieves muy distintos, sobre todo en los Andes.
- MERRA-2 es un reanálisis, no datos de estaciones; los cambios en sus datos de entrada pueden crear saltos (ver el caso de Napo).
- Cada región es un solo punto. Los de Guayas y Manaos caen sobre ciudades y ríos, así que su NDVI de MODIS describe agua y zona urbana, no cultivos ni selva.
- Los valores de Meteomatics son fotos de un modelo cada 10 días, solo en 2023–2024.
- Los modelos de humedad del suelo apenas superan a la persistencia en varias regiones y fallan en el Ártico (R² negativo en la prueba); la causa no se ha aislado.