Wind energy planning often begins with a deceptively simple question: where is the wind strongest? A new study of West Bengal, India, argues that this is not enough. Wind resources do not stop at administrative borders, and the landscapes beneath them can either support or constrain development. By combining high-resolution wind data, satellite-derived land-cover information and spatial econometric modelling, the research identifies geographically connected areas that may deserve closer investigation as future wind-energy corridors.
The analysis, published in Scientific Reports on 27 September 2026, found pronounced geographic clustering in district-level wind potential. Global Moran’s I, a measure of spatial autocorrelation, was 0.723 with p = 0.001. In practical terms, districts with relatively favourable wind conditions tended to sit near other favourable districts, while weaker areas also clustered together.
That matters because conventional regression assumes observations are independent. Here, geography itself carried information. Once the researcher explicitly modelled that dependence, the analysis suggested that tree cover and built-up land were negatively associated with wind potential, while flooded vegetation was positively associated. A pixel-level suitability analysis then delineated candidate corridors covering about 105,776 hectares across West Bengal.
From wind maps to connected landscapes
Wind-resource assessment commonly uses measurements such as average wind speed or wind power density. The new study instead built its framework around two parameters of the Weibull distribution, a statistical model widely used to describe wind-speed behaviour. The scale parameter, A, captures the characteristic magnitude of wind speed, while the shape parameter, K, describes the distribution’s form and therefore contributes information about the consistency of the wind regime.
The researcher used raster layers for these parameters at approximately 100-metre resolution. These were combined with Dynamic World Version 1 land-cover data, which are derived from Sentinel-2 satellite imagery and classify the Earth’s surface at 10-metre resolution. Annual land-cover information was resampled to align with the wind layers.
Dynamic World distinguishes categories including water, trees, grass, flooded vegetation, cropland, shrub and scrub, built areas, bare ground, and snow and ice. For the district-level statistical analysis, the researcher calculated the proportion of relevant land-cover classes in each district and examined how those characteristics related to the Weibull scale parameter.
The study included 19 district-level observations. A first-order Queen contiguity matrix defined districts as neighbours when they shared either a boundary or a vertex. This allowed the models to test not only local characteristics but also whether conditions in one district were related to those nearby.
Ordinary regression captured much of the pattern, but not the geography
The baseline ordinary least squares model explained 80.9% of the variation in the Weibull scale parameter, with an adjusted R² of 0.714. Tree cover had a coefficient of -0.027 and was statistically significant, while bare land had a coefficient of -0.259. Flooded vegetation moved in the opposite direction, with a positive coefficient of 0.586.
These coefficients should not be read as simple causal effects. A district with more tree cover may also differ in elevation, regional climate and other characteristics. The paper itself notes that some relationships may reflect where particular land-cover types are geographically concentrated rather than a direct physical effect of changing that land cover.
The spatial diagnostic made the limitation of ordinary regression clearer. Moran’s I was 0.723 and highly significant. Wind potential therefore showed strong positive spatial autocorrelation rather than a random geographic pattern.
A spatial autoregressive model improved the pseudo-R² to 0.925 and produced a spatial lag coefficient of 0.628, significant at the 1% level. This result indicates substantial similarity among neighbouring districts. Tree cover remained negatively associated with wind potential, with a coefficient of -0.0207, while flooded vegetation remained positive at 0.3935. Several other land-cover coefficients were no longer statistically significant after spatial dependence was taken into account.
The preferred model pointed to powerful unobserved regional effects
The researcher compared ordinary least squares with spatial autoregressive, spatial error and spatial Durbin models. The Spatial Error Model was ultimately selected as the preferred specification because it produced stable estimates and the most favourable information criteria, even though the Spatial Durbin Model achieved a high overall fit.
The distinction is important. In a spatial error model, geographic dependence is captured in the unexplained component of the model. The estimated spatial error coefficient, lambda, was 0.945 with a standard error of 0.035 and was highly statistically significant. Such a large coefficient suggests that unmeasured processes shared across neighbouring districts are important to the observed wind pattern.
Within this preferred model, built-up area had a coefficient of -0.011 and tree cover -0.027, both statistically significant. Flooded vegetation had a positive coefficient of 0.281. Grass, shrub and scrub, and bare land were not statistically significant in the same specification.
The findings are physically plausible as associations. Dense vegetation and urban structures increase surface roughness and can alter near-surface airflow, whereas open and seasonally flooded landscapes can present fewer obstacles. Yet the study does not establish that changing land cover would mechanically produce the estimated change in wind potential. The modelling is observational and spatial, not an experiment.
The highest-scoring areas formed corridors rather than isolated points
After estimating the district-level relationships, the researcher returned to the finer-resolution raster data. The Weibull variables were normalised and integrated with land-cover information to construct a Wind Suitability Index. Pixels at or above the 90th percentile of that index were treated as high-potential cells, allowing contiguous cells to form candidate development corridors.
The resulting corridors covered approximately 105,776 hectares. Rather than appearing uniformly within the best-performing districts, high-suitability cells formed connected geographic bands. This is the central planning contribution of the study: renewable-energy assessment may be more informative when it identifies continuous zones that can potentially share transmission, maintenance and grid-integration infrastructure.
District-level indicative estimates also varied widely. Pashchim Medinipur had the largest estimated potential at 36.30 GW, followed by Barddhaman at 28.09 GW, Bankura at 26.07 GW, Puruliya at 22.69 GW and Murshidabad at 20.68 GW. At the other end, Kolkata was estimated at just 0.05 GW, Haora at 4.13 GW and Darjiling at 8.29 GW.
Those gigawatt figures require careful interpretation. They are planning scenarios derived from representative deployment assumptions, not forecasts of electricity that will actually be generated. A technically viable wind farm depends on turbine technology, hub height, capacity factor, land access, environmental clearance, grid capacity, financing and many other conditions not resolved by a suitability map.
Why the spatial approach changes the planning question
The analysis illustrates a broader problem in infrastructure planning. Administrative averages can hide sharp local variation, while isolated site rankings can miss the fact that resources and infrastructure often operate as regional systems. A corridor approach asks whether several favourable locations form a connected development landscape.
That could matter economically. Nearby projects may be able to share grid upgrades, maintenance facilities, roads and other infrastructure. It could also matter environmentally because planning a corridor at an early stage may make it easier to identify conflicts with forests, settlements, wetlands or other sensitive land uses before individual projects are proposed piecemeal.
The strong spatial error coefficient is also a warning against treating the land-cover variables as a complete explanation. Much of the geographic structure is associated with factors that were not directly represented in the model. Atmospheric circulation, elevation, terrain, coastal effects and other regional processes may contribute to the clustering.
A planning screen, not a construction blueprint
The study has several important limitations. First, the econometric analysis used only 19 district observations. Spatial models can reveal structure in such data, but a small sample constrains statistical precision and the number of variables that can be estimated reliably.
Second, the suitability framework relies primarily on Weibull wind parameters and land-cover information. It does not substitute for site-specific wind measurements, geotechnical surveys, turbine-level engineering, transmission studies, wildlife assessment or community consultation.
Third, one of the more flexible specifications exposed a stability problem. Although the Spatial Durbin Model achieved the highest pseudo-R², its estimated spatial autoregressive parameter exceeded the theoretically admissible stability boundary. The study therefore did not use that model for its final corridor analysis, favouring the more stable Spatial Error Model.
Fourth, land cover itself changes. Urban expansion, vegetation change and altered water regimes could make today’s suitability surface less representative in future years. The framework is therefore better understood as an updateable decision-support method than as a permanent map of where turbines should go.
What the study adds
The value of the research lies less in declaring a single best place for a wind farm than in demonstrating how multiple forms of spatial evidence can be joined. The analysis moves from 100-metre wind rasters and 10-metre satellite land-cover data to district-level econometrics, then back to pixel-level suitability. That sequence allows the final map to reflect both local characteristics and the wider geographic structure detected statistically.
For West Bengal, the results point toward connected candidate areas particularly in western, southwestern and near-coastal parts of the state. For planners elsewhere, the more transferable contribution is methodological: wind-resource maps can be treated as spatial systems rather than collections of independent pixels or administrative averages.
The next step is validation. Incorporating wind speed at multiple turbine hub heights, wind power density, terrain, transmission access, protected areas, settlement buffers and field measurements would make the suitability estimates more directly useful for investment decisions. Until then, the 105,776 hectares identified here are best regarded as places to investigate more closely, not places already proven ready for construction.
Source Information
Study: Integrating spatial econometrics and remote sensing for pixel-level wind energy corridor identification: evidence from West Bengal, India
Author: Samidh Pal
Journal: Scientific Reports
Published: 27 September 2026
DOI: 10.1038/s41598-026-65979-3
