GeoAI-Based PV Suitability Screening in India Using Remote Sensing, Presence-Background Learning, and Explainable Machine Learning
Land · 29.07.2026 · 15 (8) · 1362
DOI: 10.3390/land15081362 ↗Kurz gefasst
Fachartikel in der Zeitschrift Land, online erschienen am 29.07.2026. Verfasst von Chinmay Nischal und 9 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Princess Nourah bint Abdulrahman University.
Abstract (Originalsprache)
India’s solar photovoltaic expansion requires planning tools that can identify broad zones where new ground-mounted projects may merit detailed feasibility assessment. This study develops a GeoAI screening framework that combines cleaned solar-plant reference points, open remote-sensing predictors, machine-learning models, a tabular PyTorch deep-learning model, exclusion masking, spatial validation, external inventory checks, and explainable AI. Using 2848 cleaned utility-scale solar reference points and a 17,088-sample presence-background training matrix on a 0.05-degree grid, the workflow estimated relative solar-development suitability across India. The final four-member ensemble achieved ROC-AUC = 0.834, PR-AUC = 0.591, and F1 = 0.513 under spatial holdout validation; grouped spatial cross-validation produced mean ROC-AUC = 0.776 ± 0.082, PR-AUC = 0.505 ± 0.130, and F1 = 0.406 ± 0.102. After exclusion masking, the ensemble identified 145,391 km2 of High or Very High suitability land, including 26,798 km2 in the Very High class. Under the baseline 2.0 ha MW−1 land-use assumption, developing 1–10% of the High/Very High screening area corresponded to approximately 73–727 GW of upper-bound planning-scale technical capacity. Independent World Resources Institute (WRI) and Global Energy Monitor (GEM) overlays indicated moderate spatial consistency, with 30.0–48.3% of external records falling in High or Very High classes. Sensitivity testing showed that probability cutoff, cropland treatment, and land-use assumptions strongly affect mapped extent and capacity scenarios, whereas reasonable slope-threshold changes had limited national effect. The results should therefore be interpreted as a reproducible screening and policy-prioritization surface, not as a parcel-level feasibility map.
Förderung
- Princess Nourah bint Abdulrahman University
Zitieren
APA
Nischal, C., Gupta, J., Mishra, S. K., Mir, S., Gupta, A., Singh, S., Avtar, R., Hasher, F. F. B., Gupta, M. K., & Zhran, M. (2026). GeoAI-Based PV Suitability Screening in India Using Remote Sensing, Presence-Background Learning, and Explainable Machine Learning. Land, 15(8), 1362. https://doi.org/10.3390/land15081362
Vancouver
Nischal C, Gupta J, Mishra SK, Mir S, Gupta A, Singh S, et al. GeoAI-Based PV Suitability Screening in India Using Remote Sensing, Presence-Background Learning, and Explainable Machine Learning. Land. 2026;15(8):1362. doi:10.3390/land15081362
BibTeX
@article{Nischal2026,
author = {Nischal, Chinmay and Gupta, Jagriti and Mishra, Shri Krishna and Mir, Shahnawaz and Gupta, Ankita and Singh, Saurabh and Avtar, Ram and Hasher, Fahdah Falah Ben and Gupta, Mukesh Kumar and Zhran, Mohamed},
title = {GeoAI-Based PV Suitability Screening in India Using Remote Sensing, Presence-Background Learning, and Explainable Machine Learning},
journal = {Land},
year = {2026},
volume = {15},
number = {8},
pages = {1362},
doi = {10.3390/land15081362}
}
EUVIMED