Chagas disease triatomine vector habitat suitability mapping
Landsat land cover, MODIS vegetation phenology and SRTM elevation feed species distribution models that flag where triatomine habitat is most suitable across Latin America's dry forests and scrublands. Satellite data maps the environment, not the insect itself.
Sensors
- Landsat 8/9 OLI: 30 m multispectral imagery (bands 1-7 plus SWIR) for supervised land-cover classification of dry forest, scrubland, bare soil and built-up rural areas; 16-day revisit per satellite, combined 8-day revisit with both in orbit
- MODIS Terra/Aqua (MOD13Q1/MYD13Q1): 250 m Enhanced Vegetation Index composites at 16-day intervals; captures intra-annual phenology cycles that distinguish deciduous dry forest from evergreen cover, a key discriminator in triatomine habitat models
- SRTM DEM: ~30 m (1 arc-second) global elevation model; supplies slope, aspect and altitude variables that constrain triatomine species ranges, several of which show hard upper-elevation limits around 2,000-3,000 m depending on species
- Sentinel-2 MSI: 10 m visible and 20 m red-edge/SWIR bands; used to refine peri-domestic land-cover detail (vegetable gardens, wood piles, livestock corrals) at scales where Landsat is too coarse; 5-day revisit at the equator
What the landscape tells you before the entomologist arrives
Triatomine bugs, the insect vectors of Trypanosoma cruzi, do not distribute themselves randomly. Across the Gran Chaco, the Caatinga, the Cerrado transition zones and the dry valleys of the Andes, their presence correlates with specific land-cover configurations: intact or degraded dry forest close to rural settlements, scrubland that provides refuge for sylvatic populations, and the peri-domestic clutter of wood piles, animal pens and cracked adobe walls where domestic colonies establish. None of those habitat features are invisible from orbit.
Species distribution models (SDMs) trained on published entomological survey records use environmental predictor layers derived from satellite data to estimate habitat suitability across the full geographic range of species such as Triatoma infestans, T. dimidiata, T. brasiliensis and Rhodnius prolixus. The PAHO has published surveillance frameworks that treat these models as a first-pass prioritisation tool, directing scarce field teams toward areas of highest predicted suitability rather than blanket grid surveys. Satellite data does not replace those surveys. It makes them cheaper to design.
The predictor stack: what each sensor contributes
Land-cover classification is the backbone. Landsat OLI's 30 m resolution is sufficient to distinguish the major cover classes that matter: closed dry forest, open scrubland, grassland, bare soil, irrigated agriculture and low-density built-up areas. Random forest and support vector machine classifiers trained on regional training sets routinely achieve overall accuracies above 85 percent for these broad classes in the Chaco and Caatinga, though accuracy drops in transitional zones and where cloud cover forces gap-filling from older acquisitions.
MODIS phenology layers add temporal depth. The 16-day EVI composite time series captures the seasonality signature of each cover class, which matters because triatomine habitat suitability is not static across the year. Dry-season contraction of vegetation pushes sylvatic bugs toward structures. Multi-year MODIS stacks also flag land-cover change, deforestation fronts in particular, which published studies associate with increased human-vector contact as forest-adapted species move into disturbed edge habitats.
SRTM elevation and its derivatives (slope, topographic wetness index, aspect) constrain the species-specific altitude envelopes. T. infestans, for example, has a documented range that extends into Andean inter-Andean valleys but drops off sharply above roughly 3,200 m. Sentinel-2 enters the stack at the peri-domestic scale, where 10 m resolution can resolve individual structures and their immediate surroundings in a way that 30 m Landsat cannot. It is most useful for village-scale infestation-risk scoring once a district has already been flagged by the coarser regional model.
What satellite data genuinely cannot see
Housing construction quality is the single largest determinant of domestic infestation risk, and it is largely opaque from orbit. A thatched palm roof, cracked mud-brick walls and gaps around window frames create the microhabitat T. infestans colonises. A concrete-block house on the same plot, under the same vegetation, carries far lower risk. Satellite imagery can sometimes distinguish roof material at very high resolution, but it cannot assess wall integrity, interior clutter or the presence of domestic animals sleeping indoors, all of which are documented infestation predictors in PAHO field studies.
Actual vector presence is not detectable from orbit at any resolution currently available commercially or through open constellations. SDMs produce a probability surface for habitat suitability, which correlates with, but is not equivalent to, infestation probability. A high-suitability pixel may have no bugs; a low-suitability pixel in a poorly constructed house may be heavily infested. The model is a triage instrument. Entomological ground-truth remains mandatory before any vector control intervention.
Building the species distribution model: method choices and their trade-offs
MaxEnt remains the most widely published SDM algorithm for triatomine mapping, partly because it handles presence-only occurrence records well, which is the data format most entomological surveys produce. It requires careful regularisation to avoid overfitting to the spatial bias in survey effort, a real problem when most historical records cluster around accessible towns. Ensemble approaches combining MaxEnt with Boosted Regression Trees and Random Forest reduce that bias and produce uncertainty envelopes that are honest about where the model is extrapolating beyond its training data.
Predictor selection matters as much as algorithm choice. Collinear variables (NDVI and EVI from the same composite period, for instance) inflate apparent model performance without adding information. A defensible predictor set for Chaco-region triatomines typically includes annual mean temperature, temperature seasonality (MODIS land surface temperature time series or WorldClim derivatives), dry-season EVI minimum, land-cover class, distance to forest edge, elevation and slope. Adding more variables without cross-validation against independent survey data tends to produce visually compelling but geographically overconfident outputs.
Validation against withheld survey points using AUC and the true skill statistic gives an honest read on model discrimination. Published SDMs for T. infestans in the Gran Chaco have reported AUC values in the range of 0.80 to 0.92, which is useful but not definitive. The honest position is that a well-validated model tells a programme manager where to look first, not where bugs definitely are or are not.
From suitability map to operational surveillance priority
The practical output is a district- or municipality-level prioritisation layer, typically delivered as a classified raster and a ranked attribute table that a national vector control programme can load directly into its field scheduling system. High-suitability zones with documented housing vulnerability (identified through census proxies for construction quality) move to the top of the inspection queue. Change-detection layers, derived from annual Landsat or Sentinel-2 composites, flag newly deforested areas where edge-effect contact risk is rising.
Satellize can run this analytic stack on open Landsat, Sentinel-2 and MODIS archives for any target country in Latin America, producing outputs in GeoTIFF and vector formats compatible with standard GIS environments. The approach is methodologically similar to the crop-estimation work Satellize runs for the Kingdom of Tonga: open-constellation data, calibrated against ground observations, delivered as decision-ready layers rather than raw imagery. National health ministries and international health organisations typically want the model re-run annually as land cover changes, with uncertainty bounds included in every deliverable.
Typical figures
| Land-cover classification resolution | 30 m (Landsat OLI); 10 m available with Sentinel-2 MSI for peri-domestic detail |
| Vegetation phenology resolution | 250 m, 16-day composites (MODIS MOD13Q1/MYD13Q1 EVI) |
| Elevation model resolution | ~30 m (SRTM 1 arc-second); vertical accuracy approximately ±16 m RMSE globally |
| Revisit for change detection | 8 days combined (Landsat 8+9); 5 days at equator (Sentinel-2A+2B) |
| Spectral bands used | Visible, NIR, SWIR1, SWIR2 (Landsat OLI bands 2-7); Red-edge, NIR, SWIR (Sentinel-2 MSI) |
| SDM habitat suitability output resolution | Typically 1 km (matching WorldClim/MODIS climate variables); can be sharpened to 250 m with MODIS phenology predictors |
| Archive depth | Landsat back to 1972 (TM/ETM+/OLI); MODIS back to 2000; Sentinel-2 back to 2015 |
| Minimum detectable land-cover unit | ~0.09 ha at 30 m (Landsat); ~0.01 ha at 10 m (Sentinel-2); sub-pixel mixing is a real limit in fragmented scrubland |
| Delivery formats | GeoTIFF (raster suitability surface), GeoPackage/Shapefile (district priority ranking), PDF technical report with uncertainty bounds |
| Latency for annual update | 4-8 weeks from data acquisition cutoff to validated deliverable, depending on cloud cover and ground-truth data availability |
Analytics Satellize can run
| Regional habitat suitability surface | MaxEnt or ensemble SDM (MaxEnt + BRT + Random Forest) trained on published PAHO/entomological occurrence records with Landsat land cover, MODIS EVI phenology and SRTM topography as predictors | 1 km or 250 m classified GeoTIFF with five suitability classes and per-pixel uncertainty band; country or sub-national extent |
| Dry forest and scrubland land-cover map | Supervised random forest classification of Landsat OLI multi-date composites (dry and wet season); accuracy assessment against stratified reference points | 30 m annual land-cover GeoTIFF with confusion matrix and overall accuracy report |
| Peri-domestic land-cover detail layer | Sentinel-2 10 m object-based image analysis for settlement-scale mapping of structures, bare soil, corrals and woody vegetation within 500 m of identified dwellings | GeoPackage vector layer with per-structure buffer attributes; input to village-scale infestation-risk scoring |
| Annual deforestation and edge-effect change layer | Landsat or Sentinel-2 bi-temporal change detection (post-classification comparison or spectral change vector analysis); flags newly cleared forest patches within known triatomine range | Annual change GeoTIFF and tabular summary of new edge-habitat area by administrative unit |
| District surveillance priority ranking | Spatial overlay of habitat suitability surface with census-derived housing vulnerability proxy (roof and wall material where available) and distance-to-health-facility layer | Ranked attribute table and choropleth PDF map; ready for import into national vector control scheduling systems |
| Multi-year suitability trend analysis | Time-series SDM re-runs using annual Landsat land-cover and MODIS phenology inputs from 2000 to present; identifies districts where suitability has increased due to land-use change | Trend GeoTIFF (suitability change per decade) and written interpretation for programme planning |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.