Introduction

Asbestos, an inherently fibrous silicate mineral prized for its durability and thermal resistance, remains embedded in millions of aging structures despite widespread bans1,2. When asbestos‐cement roofs degrade or sustain damage, liberated fibers pose severe health risks, including asbestosis, lung cancer, and mesothelioma3,4,5,6. These hazards are magnified in conflict or disaster settings, where structural collapse and debris generation can dramatically accelerate fiber release.

On October 7th, 2023,rocket strikes and armed incursions in Israel’s Western Negev (31° N, 34° E) wrought extensive damage to decades‐old asbestos‐cement rooftops (Fig. 1), creating an urgent need for rapid, large-area detection methods that minimize field exposure under hazardous conditions. While airborne hyperspectral imaging has shown promise for asbestos mapping in urban and geological contexts7,8,9,few studies have tackled the dual challenges of moderate (30 m) satellite resolution and constrained ground access in active conflict zones10,11.

Fig. 1
figure 1

© 2023 NV5 Geospatial Solutions Inc.)22.

Civilian areas devasted by rocket fire across the Western Negev, Israel. Image contains enhanced Copernicus data captured by Sentinel-2 on October 7th, 2023. EnMAP satellite imagery31, ENVI Remote Sensing Software (ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64],

Full size image

The launch of EnMAP, an environmental mapping hyperspectral mission, now offers spaceborne coverage at high spectral fidelity and albeit at moderate spatial resolution12,13,14,15,16. Discriminating asbestos within mixed‐material pixels requires not only robust spectral libraries but also sophisticated classification workflows capable of handling partial roof destruction, paint-sealing mandates, and heterogeneous debris.

Here, we present an end-to-end asbestos-detection pipeline that (1) anchors satellite observations with disturbance-free field spectra collected across multiple Kibbutzim and Moshavim using an ASD FieldSpec 4 High-Res spectroradiometer outfitted with our SoilPro® apparatus and a contact probe (Sup Fig. 5–14), (2) preprocesses EnMAP Level 2 imagery via MNF noise reduction, precise co-registration, and cloud/shadow masking, and (3) employs eight supervised classifiers in an iterative, multi-stage cascade, to isolate asbestos‐positive pixels under real-world war-damage conditions. We validate our approach against exhaustive ground surveys, demonstrating reliable 86% positive matches across villages and cities even amid urban complexity and security-driven coordinate restrictions. This integrative framework underscores the feasibility of orbit-based hyperspectral data for rapid, large-scale asbestos hazard mapping in high-risk environments.

Materials and methods

During a dedicated 4-month field campaign (November 2023–April 2024), we partnered with the Israel Space Agency and Ministry of Environmental Protection under Israel’s Asbestos and Harmful Dust Prevention Law (2011)17 to assemble a region-wide spectral library and establish unbiased calibration references for satellite-based mapping. Field teams visited six strategically selected villages (Sde Avraham, Sde Nitzan, Ein HaBesor, Ein HaHoresh, Nir Am, Netiv HaAsara) and two urban centers (Be’er Sheva, Sderot) to capture disturbance-free reflectance spectra from intact and partially destroyed asbestos-cement roofs (Sup Figs. 1–14). At each site, we employed an ASD FieldSpec 4 High-Res spectroradiometer (350–2500 nm) fitted with two independent fiber-optic attachments, a SoilPro® assembly for stable border-area undisturbed measurements and a standard contact probe for stable point sampling18,19. For every roof segment, five replicate spectra were recorded (integration times auto-optimized in the ViewSpecPro software20 (ViewSpec Pro version 6.2.0, ASD Inc.) as well as fore-optics set to bare fiber), with dark-current subtraction and Spectralon® 99%21 white-panel calibration repeated after every ten measurements on the same surface and before each new segment. Sampling deliberately encompassed moss and dust-covered areas, sun-weathered clean patches, and 20–25-year-old paint-sealed surfaces to capture the full continuum of end-member and mixed-pixel signatures. In total, 2714 spectra were acquired 1213 in situ ASD + SoilPro® + Contact probe runs and 1501 preliminary EnMAP pixel extracts. Government-provided coordinates of large agricultural asbestos-cement roofs (> 60 m2; e.g., barns, storage sheds, chicken coops) were reserved solely for pure end-member library construction and deliberately excluded from the final mapping process to prevent over-training or artificially inflating detection metrics (Sup Fig. 6). These exclusion sites nonetheless ensured our spectral library accurately represented the full compositional and weathering variability of asbestos-cement materials.

For satellite input, we procured EnMAP Level 2A products via the EOC EOWEB Geoportal (User group: EnMAP CAT1 Users) in BIL + Metadata format (UTM projection at scene center). These orthorectified, atmospherically corrected (land mode; ozone 330 DU; no cirrus/haze removal; automatic terrain correction; bilinear interpolation) 230-band cubes formed our primary dataset. To verify baseline radiometry, we independently processed corresponding Level 1C scenes in ENVI 6.0 using FLAASH (MODTRAN mid-latitude summer; rural aerosol model), applying as close to identical geo-correction and resampling parameters as possible22 (ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64], © 2023 NV5 Geospatial Solutions Inc.). Both L2A and L1C datasets were subjected to a Minimum Noise Fraction transform, retaining the first 15 components (capturing > 99% of scene variance) to suppress sensor noise while preserving critical absorption features. Each scene was co-registered to GPS control points (RMSE < 0.3 px), then masked for clouds and shadows via combined NDVI < 0.10 and near-IR brightness > 0.30 reflectance thresholds. Finally, we spectrally resampled the corrected cubes to our field library wavelengths, ensuring a one-to-one match between satellite and ground spectra before classification.

Building on our meticulously curated spectral library and pre-processed imagery, we then applied eight complementary supervised classifiers; Linear Spectral Unmixing (LSU)23, Support Vector Machine (SVM) with RBF kernel (C = 1, γ = 1/nbands)24,25, Spectral Angle Mapper (SAM) with a 0.10 rad threshold23,26,27, Adaptive Coherence Estimator (ACE)27, Mahalanobis Distance (3 σ cutoff)28, Maximum Likelihood Classification (MLC) under equal-prior Gaussian assumptions24,26,27, Spectral Information Divergence (SID)23,26, and Matched Filtering (MF) tuned to pure-asbestos end-members24,26, each executed independently in ENVI 6.022. We chose this ensemble to span physically based unmixing (LSU, MF), statistical discrimination (Mahalanobis, ACE, SID), angular similarity (SAM), and machine-learning (SVM, MLC), ensuring that diverse aspects of the asbestos spectral signature were probed. To fuse their outputs into a robust detection mask, we instituted an iterative consensus cascade: only pixels flagged positive by at least six of the eight classifiers were retained; these survivors were then passed through each method again to purge marginal outliers. Where classifier outputs conflicted i.e., a pixel met the six-of-eight rule but showed divergent raw scores we inspected its full spectral profile and compared it to our high-fidelity library spectra, adjusting decision thresholds in ENVI’s Classification Workflow Designer to eliminate any residual false positives without sacrificing recall (ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64], © 2023 NV5 Geospatial Solutions Inc.)22.

Building on our meticulously tuned classifiers, we then conducted independent ground-truth surveys from (May 2024 through August 2024) in ten villages (Kfar Aza, Nir Oz, Nahal Oz, Re’im, Be’eri, Sde Nitzan, Urim, Nirim, En Hashlosha, Kerem Shalom) and two cities (Netivot, Sderot). Field teams, guided solely by the 823 newly identified candidate pixels on our maps, visited each rooftop location, whether intact, partially destroyed by rocket fire, or burned and recorded the presence or absence of asbestos-cement material (Figs. 2 and 3). This process uncovered 823 discrete rooftop detections not previously catalogued, and the resulting comparison of field observations against our final binary mask yielded an 86% positive match rate (Fig. 3). Accuracy metrics (overall accuracy, precision, recall, F1-score, Cohen’s κ) were calculated directly in ENVI’s Accuracy Assessment module from the confusion matrix 22. To guarantee data integrity, all field logs underwent double-entry verification, photographic timestamps were matched to GPS metadata, and coordinate accuracy was confirmed to within 1 m using differential GPS. These rigorous, blind validation steps confirm that our workflow’s performance reflects genuine predictive power in complex, conflict-affected settings, untainted by over-training or confirmation bias.

Fig. 2
figure 2

© 2023 NV5 Geospatial Solutions Inc.)22.

Gound truth and EnMap findings are overlayed on an aerial photo taken by UAV. Green tiles reveal positive classification pixels of asbestos based cement rooftops, while the red tiles represent the actual structure of all roofing that is contcated by the postive identification. *The aerial photograph was captured before the start of the war. EnMAP satellite imagery 31, Base map 33,ENVI Remote Sensing Software(ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64],

Full size image
Fig. 3
figure 3

© 2023 NV5 Geospatial Solutions Inc.)22.

Selected examples from confirmed Asbestos based cement rooftops and non-Asbestos targets nearby, from EnMAP data. EnMAP satellite imagery31, ENVI Remote Sensing Software (ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64],

Full size image

Results

We assessed the accuracy of seven supervised classifiers against 823 blind ground‐truth rooftop detections (May–August 2024) (Fig. 3). (Table 1) and (Table S1) summarize key performance metrics such as, overall accuracy, precision, recall, F1‐score, and Cohen’s κ, for each method against the 823 detections that were located across the ten villages and two cities.

Table 1 Summary of key performance metrics overall accuracy, precision, recall, F1‐score, and Cohen’s κ, for each method against the 823 detections that were located across the ten villages and two cities.
Full size table

Key Findings.

  • Top performers: ACE led with 91.4% accuracy (κ = 0.87), followed by SID (90.1%, κ = 0.85) and SVM (89.2%, κ = 0.83).

  • Lower bound: LSU (84.5%, κ = 0.79) was most challenged by mixed–pixel scenarios.

  • Balanced sensitivity: Recall ranged from 86.1 to 92.7%, confirming strong true-positive detection, while precision (82.3–89.9%) reflected effective false-positive suppression.

Although ACE alone reached 91.4% accuracy, this peak performance emerges only after a meticulous, multi-stage workflow. Each classifier was first applied independently; their outputs were then woven together in an ensemble filter requiring six-of-eight agreement. Initial runs produced over 2300 candidate pixels; through successive consensus passes and manual inspection, wherein each surviving pixel’s spectrum was directly compared to our high-fidelity reference library, false positives were progressively culled. This iterative curation reduced the candidate set to 823 final detections and underscores the necessity of human‐in‐the‐loop validation to mitigate errors that single methods cannot resolve.

We observed higher false positives in areas with concrete rubble or paint‐sealed fragments, particularly in dense urban blocks where 30 m pixels spanned mixed materials. SAM and LSU were most affected, while ACE and SID remained robust. Rural Kibbutzim and Moshavim, with larger homogeneous roofs, consistently achieved > 92% match rates, highlighting the influence of spatial context on mixed-pixel confusion.

Our ensemble‐driven approach reaches or exceeds accuracy benchmarks (80–90%) from non‐conflict asbestos‐mapping studies but uniquely maintains high performance under war‐damage conditions29,30. The manual, spectrum‐by‐spectrum verification step is critical for transferring such pipelines to other high‐risk, complex settings.

Discussion

This work establishes a coherent, reproducible pipeline for mapping asbestos‐cement roofs under the most challenging conditions, active conflict, debris fields, and restricted ground access, by combining disturbance-free field calibration, rigorous preprocessing of EnMAP data14,31,32, and a multi-stage ensemble of eight classifiers with human-in-the-loop verification. Over a 4-month campaign (Nov 2023–Apr 2024), we collected 2714 spectra across six villages (Sde Avraham, Sde Nitzan, Ein HaBesor, Ein HaHoresh, Nir Am, Netiv HaAsara) and two cities (Be’er Sheva, Sderot), deliberately sampling intact roofs, moss-and-dust-covered patches, sun-weathered areas, and 20–25-year-old painted seals. We collected five replicate spectra per measurement per roof segment, with white referencing done on the Spectralon21 for recalibration after every ten scans, ensuring high-quality, disturbance-free field measurements across all villages and cities. These spectra formed a high-fidelity library that underpinned every classifier, supplying pure end-members for LSU and MF, training sets for SVM and MLC, and reference signatures for SAM, ACE, Mahalanobis, and SID.

Each classifier was run independently in ENVI 6.022, then their outputs were woven together in an iterative consensus cascade requiring six-of-eight agreement. Starting from over 2,300 candidate pixels, successive consensus passes eliminated conflicting detections, and the remaining pixels underwent manual spectrum-by-spectrum curation against our reference library. This painstaking workflow distilled the set to 823 final detections and yielded an 86% positive match rate against truly blind ground-truth surveys conducted (May–Aug 2024) in ten villages (including Kfar Aza, Nir Oz, Nahal Oz, Re’im, Be’eri, Sde Nitzan, Urim, Nirim, En Hashlosha, Kerem Shalom) and two cities (Fig. 3). The fact that ACE reached 91.4% accuracy and κ = 0.87 only after this ensemble-plus-manual process and that LSU alone peaked at 84.5%, underscores the necessity of both automated consensus and human oversight. We will continue to expand and improve our study as we begin to map the entire country of Israel, furthering the techniques and validation of this ongoing research.

Despite these strong results, our reliance on 30 m EnMAP14,31,32 pixels inevitably blend heterogeneous materials, roof fragments, adjacent concrete rubble, and painted surfaces, giving rise to mixed-pixel errors. SAM and LSU, which depend on strict spectral angles or linear unmixing, suffered the falsest positives in dense urban blocks where rocket-damaged debris mimics asbestos features; ACE and SID proved more robust by leveraging statistical coherence and entropy-based divergence. Rural Kibbutzim and Moshavim, characterized by larger, homogeneous roofs, consistently achieved > 92% match rates, highlighting how spatial context and pixel scale influence detection performance.

We also confronted the challenge of paint sealing: aged, weathered paint alters the rooftop’s spectral signature and initially produced false negatives. By explicitly sampling painted surfaces and including them in our library, we dramatically improved classifier recall. However, long-term weathering and variable debris coverage remain obstacles that may benefit from deep-learning approaches capable of learning complex, non-linear spectral-spatial patterns without manual threshold tuning.

Finally, while our pipeline proved effective in Israel’s semi-arid Western Negev, where sparse vegetation and minimal lichen colonization facilitate clear roof spectra, transferring it to temperate or humid climates will require building local spectral libraries that capture additional confounders such as lichen, moss, and tree canopy interference. Integrating UAV-borne hyperspectral and LiDAR data for sub-meter roof delineation, leveraging time-series EnMAP14,31,32 acquisitions to distinguish transient damage from persistent asbestos signatures, and incorporating deep-learning classifiers offer promising avenues to further enhance accuracy and operational efficiency.

From a public health perspective, this ensemble workflow delivers rapid, large-scale asbestos‐hazard maps, critical for prioritizing decontamination, coordinating safe field inspections, and informing policy under crisis conditions (Fig. 4). By transparently documenting our sampling schedules, preprocessing parameters, classifier settings, and validation protocols, we provide a template that other regions can adapt and refine, enabling timely hazard surveillance where traditional surveys are impractical or unsafe (Fig. 4).

Fig. 4
figure 4

© 2023 NV5 Geospatial Solutions Inc.)22.

Heat-map of classified asbestos-cement rooftops (colour scale indicates density) overlaid on county boundaries. Blue line: administrative boundary of Western Negev counties. EnMAP satellite imagery31, ENVI Remote Sensing Software (ENVI 6.0, Build 11/15/2023 [Microsoft Windows 64/64],

Full size image

Conclusions

Using EnMAP satellite imagery14,31,32 and a rigorous, human-in-the-loop ensemble workflow, we have demonstrated reliable, large-scale mapping of asbestos-cement rooftops under active conflict conditions. By collecting five replicate spectra per roof segment with a Spectralon®21 recalibration every ten scans to build a disturbance-free spectral library and then applying eight complementary classifiers in a six-of-eight consensus cascade followed by manual spectrum-by-spectrum verification, we distilled over 2,300 initial candidates down to 823 validated detections. This process yielded an overall 86% positive match rate and up to 91.4% accuracy (ACE; κ = 0.87), even amid rocket-damaged debris and paint-sealed surfaces.

Although our pipeline proved robust in Israel’s semi-arid Western Negev, where sparse vegetation and minimal lichen cover simplify roof spectra, transferring it to temperate or humid climates will require building local libraries that capture additional confounders such as moss, lichen, and tree canopy interference. The 30 m spatial resolution of EnMAP14,31,32 imposes mixed-pixel challenges, especially in dense urban blocks, but integrating UAV-borne hyperspectral and LiDAR data for sub-meter roof delineation, leveraging repeated time-series acquisitions, and adopting deep-learning classifiers to automate pixel-level interpretation, all offer promising paths to enhance spatial specificity and reduce manual effort.

Ultimately, this integrative framework, anchored in robust field calibration, multi-algorithm spectral filtering, and exhaustive blind validation, provides a scalable blueprint for environmental health surveillance in crisis or inaccessible settings. By equipping responders and policymakers with rapid, high‐accuracy asbestos-hazard maps, it can guide targeted remediation, reduce exposure risks, and inform long-term public‐health interventions wherever traditional surveys are impractical or unsafe.