Dynamic Habitat Index (DHI)

• Dynamic Habitat Index (DHI) is a remote sensing-derived measure summarizing intra-annual ecosystem functioning through annual mean, minimum, and maximum radiometric indices.
• It is applied in species distribution models to compare continuous functional metrics with categorical land-use/land-cover predictors.
• DHI-based models yield broader, spatially neutral suitability maps that mitigate spatial biases inherent in distance-based LULC predictors.

Searching arXiv for the specified paper and any directly relevant DHI background sources. The Dynamic Habitat Index (DHI) is a remote sensing-based measure of habitat productivity and variability that was compared with land-use/land-cover (LULC) metrics for species distribution modelling (SDM) in Île-de-France in "Modelling species distributions using remote sensing predictors: Comparing Dynamic Habitat Index and LULC" (Oliveira et al., 18 Sep 2025). In that study, DHI is presented as a compact summary of intra-annual ecosystem functioning derived from time series of radiometric indices, classically defined through cumulative productivity, minimum cover, and variability, but operationalized there as annual mean, annual minimum, and annual maximum computed from Sentinel-2 observations for 2018 (Oliveira et al., 18 Sep 2025). The study’s central result is that DHI-based and LULC-based SDMs had similar reliability by the Continuous Boyce Index, but diverged after binarization because DHI-based suitability surfaces were broader and more spatially neutral, whereas LULC-based models were more spatially constrained by distance effects (Oliveira et al., 18 Sep 2025).

1. Definition and conceptual basis

The study defines the Dynamic Habitat Index as a compact summary of intra-annual ecosystem functioning from remote sensing, classically built from a one-year time series $x_1, x_2, \ldots, x_T$ of a radiometric index (Oliveira et al., 18 Sep 2025). In its canonical form, the three components are cumulative or annual productivity, minimum cover or greenness, and variability or seasonality. These are given as:

• $P = \sum_{t=1}^{T} x_t$
• $M = \min_t x_t$
• $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$ (Oliveira et al., 18 Sep 2025)

In the Île-de-France study, the operationalization departs from the canonical cumulative/CV formulation and instead uses statistics described as robust and straightforward to compute with Sentinel-2 (Oliveira et al., 18 Sep 2025). For each pixel, and separately for each index, the authors computed:

• $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$
• $DHI_{\mathrm{min}}(x) = \min_t x_t$
• $DHI_{\mathrm{max}}(x) = \max_t x_t$ (Oliveira et al., 18 Sep 2025)

Here, $x_t$ denotes the index value from valid Sentinel-2 Level 2A observations in 2018, and $T$ is the number of valid acquisitions for that pixel in the year (Oliveira et al., 18 Sep 2025). DHI is therefore treated as a three-component vector derived separately for each index rather than as a single composite score. No additional normalization of DHI components is described; standardization enters later through model selection and evaluation (Oliveira et al., 18 Sep 2025).

This formulation places DHI in the category of continuous, temporally integrated habitat predictors. A plausible implication is that its informational content differs fundamentally from categorical habitat descriptors because it summarizes ecosystem functioning directly from sensor-derived radiometric behavior rather than from human-defined land-cover classes.

2. Remote-sensing construction of DHI in the Île-de-France study

The DHI implementation in the study was based on Sentinel-2 Level 2A surface reflectance imagery for calendar year 2018, distributed by Theia, with atmospheric correction and cloud masking performed using the MAJA processor (Oliveira et al., 18 Sep 2025). Indices were computed at 10 m native resolution and then upscaled to 50 m to match occurrence data accuracy (Oliveira et al., 18 Sep 2025). For each index, all valid observations across the 2018 time series were aggregated to compute $DHI_{\mathrm{mean}}$, $P = \sum_{t=1}^{T} x_t$0, and $P = \sum_{t=1}^{T} x_t$1; specific compositing windows and gap-filling procedures are not detailed (Oliveira et al., 18 Sep 2025). Orfeo ToolBox v7.2.0 was used to compute the index and DHI component layers (Oliveira et al., 18 Sep 2025).

Eighteen spectral or radiometric indices spanning vegetation, soil, and water were computed, each summarized by the three DHI statistics (Oliveira et al., 18 Sep 2025). The indices listed include NDVI, TSAVI, GEMI, RVI, MNDWI, MSAVI, MSAVI2, NDWI/NDWI2, IPVI, SAVI, TNDVI, NDTI, BI, CI, and ISU, among others (Oliveira et al., 18 Sep 2025). EVI, fPAR, LAI, and NPP proxies were not used (Oliveira et al., 18 Sep 2025). The most influential remote-sensing predictors across species derived from NDVI and TSAVI, with $P = \sum_{t=1}^{T} x_t$2 described as overwhelmingly dominant (Oliveira et al., 18 Sep 2025).

Several formulas are explicitly given for the remote-sensing indices used in the models. NDVI is defined as

$P = \sum_{t=1}^{T} x_t$3

with Sentinel-2 bands $P = \sum_{t=1}^{T} x_t$4 and $P = \sum_{t=1}^{T} x_t$5 (Oliveira et al., 18 Sep 2025). RVI is

$P = \sum_{t=1}^{T} x_t$6

(Oliveira et al., 18 Sep 2025). TSAVI is

$P = \sum_{t=1}^{T} x_t$7

where $P = \sum_{t=1}^{T} x_t$8 and $P = \sum_{t=1}^{T} x_t$9 are the slope and intercept of the soil line (Oliveira et al., 18 Sep 2025). GEMI is

$M = \min_t x_t$0

with

$M = \min_t x_t$1

(Oliveira et al., 18 Sep 2025). MNDWI is

$M = \min_t x_t$2

with $M = \min_t x_t$3 and $M = \min_t x_t$4 commonly $M = \min_t x_t$5 or $M = \min_t x_t$6 (Oliveira et al., 18 Sep 2025).

The study also notes that EVI was not computed. Its standard form is given only hypothetically and not used in the reported models (Oliveira et al., 18 Sep 2025). This distinction matters because the paper’s conclusions concern a DHI formulation grounded in the indices actually included, not in a broader class of vegetation-function products.

3. Comparator framework: LULC descriptors and species distribution modelling

The comparison to DHI was performed using continuous predictors derived from a categorical LULC map rather than by using categorical land-cover labels directly (Oliveira et al., 18 Sep 2025). LULC was obtained from a 2018 Sentinel-2 time series via a supervised Random Forest classification, specifically the OSO–CESBIO product, and aggregated into 11 classes: Urban area, Roads, Winter crops, Summer crops, Grassland, Orchards, Vineyards, Deciduous forest, Coniferous forest, Woody moorlands, and Water (Oliveira et al., 18 Sep 2025).

Two classes of LULC-derived predictors were used. The first consisted of area metrics, that is, the proportion or area of each class in local neighborhoods as provided by FRAGSTATS (Oliveira et al., 18 Sep 2025). The second consisted of distance-to-class metrics. For each class $M = \min_t x_t$7, the Euclidean distance from pixel $M = \min_t x_t$8 to the nearest pixel of class $M = \min_t x_t$9 was defined as

$$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$0

where $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$1 is the set of pixels belonging to class $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$2 (Oliveira et al., 18 Sep 2025). Distances were computed at the 50 m analysis grid and used directly as continuous predictors; no multi-scale buffers or transformations beyond the FRAGSTATS computation were reported (Oliveira et al., 18 Sep 2025).

The SDM experiment was conducted over Île-de-France, approximately $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$3, described as temperate landscapes dominated by urban and intensive agriculture (Oliveira et al., 18 Sep 2025). Species data covered eleven bird, amphibian, and mammal species from opportunistic occurrences in the Cettia ÎDF citizen database (GeoNat’ÎdF) from 2015 to 2020 (Oliveira et al., 18 Sep 2025). The species were:

• Birds: Athene noctua, Anthus pratensis, Pyrrhula pyrrhula, Sylvia curruca
• Amphibians: Bufo bufo, Hyla arborea, Ichthyosaura alpestris, Lissotriton vulgaris, Triturus cristatus
• Mammals: Eptesicus serotinus, Meles meles (Oliveira et al., 18 Sep 2025)

Occurrence records were spatially thinned to 50 m, with one presence per $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$4 m pixel, to reduce spatial sorting bias and match environmental grid resolution (Oliveira et al., 18 Sep 2025). The presence-pseudoabsence design used five pseudo-absence sets per species, randomly sampled across the background, with the number of pseudo-absences equal to the number of presences and equal weighting of presences and pseudo-absences (Oliveira et al., 18 Sep 2025).

Predictor selection differed between the remote-sensing and LULC branches. The remote-sensing set comprised 10 variables selected once and reused for all species after collinearity screening with Pearson $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$5 using removeCollinearity in virtualspecies (Oliveira et al., 18 Sep 2025). The LULC set used species-specific variable importance from biomod2, retaining the 10 top predictors per species (Oliveira et al., 18 Sep 2025). Modelling employed nine algorithms—GLM, GAM, MARS, ANN, FDA, CTA, GBM, RF, and MAXENT—with three calibrations per algorithm, a 70% train and 30% test split, and an ensemble defined as the arithmetic mean of suitability predictions from models with AUC $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$6 (Oliveira et al., 18 Sep 2025). The platform was biomod2 v3.5-3 in R (Oliveira et al., 18 Sep 2025).

4. Evaluation framework and quantitative comparison

Model performance was assessed using the Continuous Boyce Index (CBI) and a calibrated AUC denoted AUCc (Oliveira et al., 18 Sep 2025). CBI is defined in the study as a presence-only reliability metric based on a predicted-to-expected ratio across suitability bins. If bins $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$7 partition the suitability range, with $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$8 the frequency of presences in bin $$S = \mathrm{CV}(x_t) = \mathrm{sd}(x_t)/\mathrm{mean}(x_t)$$9 and $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$0 the expected frequency under random use proportional to area of bin $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$1, then

$DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$2

and CBI is computed as the Spearman rank correlation between bin mid-suitability and $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$3 across bins, ranging from $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$4 to $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$5 (Oliveira et al., 18 Sep 2025). The study implemented CBI as in Hirzel et al. (2006) (Oliveira et al., 18 Sep 2025).

AUCc is described as presence-absence discrimination adjusted for spatial sorting bias by calibrating against a geographic null model (Oliveira et al., 18 Sep 2025). Operationally,

$DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$6

where $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$7 arises from a null model driven by geographic distance to training presences (Oliveira et al., 18 Sep 2025). The reported AUCc values ranged from 0.27 to 0.48, indicating low discrimination beyond geographic bias, with LULC consistently higher than remote sensing (Oliveira et al., 18 Sep 2025).

For thresholding, the study used the 10th percentile of suitability at presence locations. If the sorted suitability scores at presences are $DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$8, the threshold is

$DHI_{\mathrm{mean}}(x) = (1/T)\sum_{t=1}^{T} x_t$9

and pixels with suitability $DHI_{\mathrm{min}}(x) = \min_t x_t$0 are classified as suitable (Oliveira et al., 18 Sep 2025).

Overlap between remote-sensing and LULC niches was evaluated on continuous suitability surfaces normalized to sum to 1 across the domain, using Schoener’s $DHI_{\mathrm{min}}(x) = \min_t x_t$1, Warren’s $DHI_{\mathrm{min}}(x) = \min_t x_t$2, and Spearman rank correlation (Oliveira et al., 18 Sep 2025). The formulas reported are:

$DHI_{\mathrm{min}}(x) = \min_t x_t$3

$DHI_{\mathrm{min}}(x) = \min_t x_t$4

$DHI_{\mathrm{min}}(x) = \min_t x_t$5

where $DHI_{\mathrm{min}}(x) = \min_t x_t$6 is the rank difference between paired suitability values across pixels (Oliveira et al., 18 Sep 2025).

The main comparative results are concise but consequential. For CBI, both remote-sensing and LULC ensembles achieved high reliability, with values close to 1 for most species; Eptesicus serotinus was the main exception with weaker CBI (Oliveira et al., 18 Sep 2025). For AUCc, LULC models consistently outperformed remote-sensing models after calibrating out spatial sorting bias, but all values remained low, indicating modest discrimination beyond geography, particularly for the remote-sensing branch (Oliveira et al., 18 Sep 2025).

Before binarization, the continuous suitability surfaces were broadly concordant. Schoener’s $DHI_{\mathrm{min}}(x) = \min_t x_t$7 ranged from 0.722 to 0.827, Warren’s $DHI_{\mathrm{min}}(x) = \min_t x_t$8 from 0.934 to 0.972, and Spearman $DHI_{\mathrm{min}}(x) = \min_t x_t$9 from 0.461 to 0.757 across species (Oliveira et al., 18 Sep 2025). After binarization, however, the niches diverged substantially. The projected overlap onto LULC niches was $DHI_{\mathrm{max}}(x) = \max_t x_t$0 on average, whereas the projected overlap onto remote-sensing niches was $DHI_{\mathrm{max}}(x) = \max_t x_t$1 on average (Oliveira et al., 18 Sep 2025). The ratio of off-overlap, defined in the paper as $DHI_{\mathrm{max}}(x) = \max_t x_t$2, was $DHI_{\mathrm{max}}(x) = \max_t x_t$3 on average, with remote-sensing off-overlap generally much larger; Hyla arborea exhibited approximately $DHI_{\mathrm{max}}(x) = \max_t x_t$4 (Oliveira et al., 18 Sep 2025).

5. Spatial behavior, bias structure, and predictor importance

A central finding of the study is that the two predictor families impose different spatial behavior on suitability maps (Oliveira et al., 18 Sep 2025). LULC-based models showed a pronounced distance effect: suitability decreased as distance to key LULC classes increased (Oliveira et al., 18 Sep 2025). In contrast, remote-sensing-based models using continuous DHI predictors were not affected by distance-to-class or geographic sampling bias in the same way; their suitable areas extended beyond observed clusters and often appeared as scattered isolated pixels (Oliveira et al., 18 Sep 2025).

This contrast is visible most clearly after thresholding. The study states that remote-sensing niches were broader and more dispersed, while LULC niches were more spatially constrained and clustered around core habitats (Oliveira et al., 18 Sep 2025). The ratio $DHI_{\mathrm{max}}(x) = \max_t x_t$5 tended to decrease with mean inter-occurrence distance for most species, meaning that broader observed distributions usually yielded less disparity between remote-sensing and LULC off-overlap, although exceptions were reported for Athene noctua, Lissotriton vulgaris, and Triturus cristatus (Oliveira et al., 18 Sep 2025).

Feature-importance analysis identifies the specific DHI components responsible for much of the remote-sensing model behavior. Among DHI predictors, $DHI_{\mathrm{max}}(x) = \max_t x_t$6 was the most influential predictor across taxa, with permutation importance from 0.25 to 0.98 (Oliveira et al., 18 Sep 2025). In several species it alone could suffice: Bufo bufo, Ichthyosaura alpestris, Triturus cristatus, Lissotriton vulgaris, and Pyrrhula pyrrhula (Oliveira et al., 18 Sep 2025). For Athene noctua, $DHI_{\mathrm{max}}(x) = \max_t x_t$7 was the sole sufficient variable (Oliveira et al., 18 Sep 2025). Other contributing remote-sensing variables included $DHI_{\mathrm{max}}(x) = \max_t x_t$8, $DHI_{\mathrm{max}}(x) = \max_t x_t$9, $x_t$0, $x_t$1, $x_t$2, and $x_t$3 (Oliveira et al., 18 Sep 2025).

By contrast, the LULC branch exhibited no single dominant landscape metric, with maximum importance not exceeding 0.41 (Oliveira et al., 18 Sep 2025). For every species, at least one distance-to-class variable ranked among the top three predictors; examples given include distance to deciduous forest, urban areas, orchards, and water (Oliveira et al., 18 Sep 2025). Area metrics were also important, but the paper states that they contributed less to spatial constraint than the distance variables (Oliveira et al., 18 Sep 2025).

This pattern suggests that the contrast between DHI and LULC is not merely one of data source, but of predictor geometry. A plausible implication is that continuous functional summaries derived from radiometric time series can produce suitability fields less tightly tethered to mapped habitat edges than metrics defined directly by Euclidean proximity to categorical classes.

6. Interpretation, limitations, and methodological implications

The study identifies conditions under which DHI is preferable. It is recommended for regional-scale SDM when continuous, temporally integrated habitat proxies can mitigate biases caused by uneven sampling and avoid artificial spatial constraints imposed by categorical LULC classes and their distances (Oliveira et al., 18 Sep 2025). It is also described as suitable when spatially neutral, sensor-derived measures of productivity and greenness are desired without reliance on human-defined classes (Oliveira et al., 18 Sep 2025).

At the same time, the paper emphasizes several limitations and caveats. The temporal window was one year only, specifically 2018; DHI sensitivity to anomalous years or interannual variability is described as real, and the authors note that incorporating multiple years and the variability term $x_t$4 as coefficient of variation may better capture seasonality (Oliveira et al., 18 Sep 2025). Cloud or snow contamination and variable scene counts per pixel can affect DHI robustness; although MAJA masks clouds, explicit gap-filling was not detailed (Oliveira et al., 18 Sep 2025). The sensor and metric choice also matters: while NDVI-based DHI dominated in this study, the paper notes that other functional products such as fPAR, LAI, and GPP can outperform NDVI or EVI for biodiversity in some contexts (Oliveira et al., 18 Sep 2025).

The study also notes ecological blind spots shared, in different ways, by both predictor families. Habitat types not captured by productivity proxies or by coarse LULC typologies remain difficult to model; the paper gives the example of very small ponds critical to amphibians (Oliveira et al., 18 Sep 2025). Remote-sensing maps tended to predict scattered suitable pixels, which the authors identify as possible overestimation, whereas LULC maps underpredicted away from observed clusters because of distance effects (Oliveira et al., 18 Sep 2025). For that reason, combining both predictor families is presented as a way to mitigate complementary biases (Oliveira et al., 18 Sep 2025).

Generalizability is treated cautiously. The approach is described as transferable to other regions and taxa, but performance depends on species ecology, occurrence bias, the ecological fit of the LULC typology, and the temporal and spatial resolution of the remote-sensing data (Oliveira et al., 18 Sep 2025). This suggests that DHI should not be understood as a universal replacement for LULC, but as a specific predictor family whose advantages emerge most clearly under spatial bias and categorical-distance constraints.

7. Position of DHI in regional SDM according to the study

The integrative conclusion of the paper characterizes the Dynamic Habitat Index as a compact, multi-component summary of intra-annual ecosystem functioning from remote sensing (Oliveira et al., 18 Sep 2025). In the study’s implementation, it was operationalized with Sentinel-2 as three per-index statistics—annual mean, annual minimum, and annual maximum—computed for 18 vegetation, soil, and water indices and used as continuous predictors in SDMs (Oliveira et al., 18 Sep 2025). Compared with LULC metrics based on area and Euclidean distance to classes, DHI-based models achieved similar reliability, with CBI near 1, but slightly lower calibrated discrimination after removing geographic sorting bias (Oliveira et al., 18 Sep 2025).

The decisive difference lay in the geometry of predicted niches. DHI predictors yielded suitability surfaces described as spatially neutral and not decaying with distance from observed habitats, thereby mitigating the geographic bias inherent to distance-based LULC predictors (Oliveira et al., 18 Sep 2025). After binarization, remote-sensing or DHI niches were consistently broader and more dispersed, while LULC niches were more constrained and clustered, with quantitative overlap metrics showing that remote-sensing off-overlap was on average about 2.6 times larger than LULC off-overlap, with $x_t$5, $x_t$6, and $x_t$7 (Oliveira et al., 18 Sep 2025).

The dominance of $x_t$8 across species, and of $x_t$9 for Athene noctua, underscores annual productivity and greenness as the principal functional drivers at the regional scale in this dataset (Oliveira et al., 18 Sep 2025). The paper therefore supports DHI as a spatially neutral, temporally informed alternative to LULC for regional SDM, particularly when occurrence data are spatially biased and categorical habitat distances induce artificial constraints (Oliveira et al., 18 Sep 2025). It simultaneously concludes that both predictor families are complementary: LULC captures structural composition and connectivity, while DHI captures functional dynamics and seasonality (Oliveira et al., 18 Sep 2025). A plausible implication is that future regional SDM workflows may benefit most from joint formulations that preserve this complementarity rather than privileging either predictor family in isolation.