Laxman Bokati
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Laxman Bokati has a total of 10 co-authors. Here are the top co-authors by number of shared publications:
- Saurav Kumar
Arizona State Univ
5 shared publications - Anil Somenahally
Texas A&M AgriLife Research, Texas A&M University, Overton, USA
3 shared publications - Aaron Andrew Velasco
University of Texas at El Paso, 500 W. University Ave, El Paso, TX, United States [US
2 shared publications - Vladik Kreinovich
Department of Computer Science, University of Texas at El Paso, 500 W. University, El Paso, TX 79968, USA
2 shared publications - Atharva Gupta1 shared publication
- Javad Robatjazi
Texas A&M AgriLife Research and Extension Center at Overton, 1710 FM 3053 N, Overton, TX, 75684, USA
1 shared publication
Latest Publications
- Open Access
- 5 September 2026
- Published by Elsevier in SSRN Electronic Journal
On Earth Observation Foundation Model Encoding Ecological Disturbance? A Zero-Shot Geometric Probe of Prithvi-EO-2.0
Earth Observation (EO) foundation models are widely deployed as frozen feature extractors, yet whether pretrained representations encode specific physical landscape processes remains under-examined. We use a parameter-free, zero-shot geometric probe to understand how wildfire impacts are encoded. The probe trains no classifier and has no learned parameters; it operates purely as cosine geometry on embedding difference vectors, so any structure it reveals resides in the representation itself. We use pretrained embeddings from Prithvi-EO-2.0-300M, a Vision Transformer trained on multispectral imagery, with no fine-tuning. From 50 globally distributed wildfire sites spanning six regions, we extract paired pre-fire and post-fire patch embeddings and construct site-level change vectors in the 1024-dimensional embedding space. Across all 1,225 site pairs, pairwise cosine similarity averages 0.813. To account for anisotropy in the embedding space, we used a permutation null that preserves the anisotropy of the change vectors while removing genuine cross-site differences. The null similarity was 0.761, 17.9 standard deviations below the observed 0.813, indicating that wildfire is encoded as a consistent direction across diverse biomes rather than an artifact of the embedding geometry. As a control condition, a ±2-year window shows near-zero embedding shift in non-fire years, establishing specificity of the fire-year signal against interannual background variability. A wildfire-versus-logging selectivity check (cosine similarity 0.441 to 0.717) indicates that the probe may tap a general canopy-loss axis rather than fire-specific physics. By measuring geometry rather than training a classifier, this approach offers a model-agnostic way to evaluate what EO foundation model representations encode.
- 5 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessOn Earth Observation Foundation Model Encoding Ecological Disturbance? A Zero-Shot Geometric Probe of Prithvi-EO-2.0
Earth Observation (EO) foundation models are widely deployed as frozen feature extractors, yet whether pretrained representations encode specific physical landscape processes remains under-examined. We use a parameter-free, zero-shot geometric probe to understand how wildfire impacts are encoded. The probe trains no classifier and has no learned parameters; it operates purely as cosine geometry on embedding difference vectors, so any structure it reveals resides in the representation itself. We use pretrained embeddings from Prithvi-EO-2.0-300M, a Vision Transformer trained on multispectral imagery, with no fine-tuning. From 50 globally distributed wildfire sites spanning six regions, we extract paired pre-fire and post-fire patch embeddings and construct site-level change vectors in the 1024-dimensional embedding space. Across all 1,225 site pairs, pairwise cosine similarity averages 0.813. To account for anisotropy in the embedding space, we used a permutation null that preserves the anisotropy of the change vectors while removing genuine cross-site differences. The null similarity was 0.761, 17.9 standard deviations below the observed 0.813, indicating that wildfire is encoded as a consistent direction across diverse biomes rather than an artifact of the embedding geometry. As a control condition, a ±2-year window shows near-zero embedding shift in non-fire years, establishing specificity of the fire-year signal against interannual background variability. A wildfire-versus-logging selectivity check (cosine similarity 0.441 to 0.717) indicates that the probe may tap a general canopy-loss axis rather than fire-specific physics. By measuring geometry rather than training a classifier, this approach offers a model-agnostic way to evaluate what EO foundation model representations encode.
- 22 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 198
Ensemble streamflow forecasting with diverse loss functions
- 22 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 198
Ensemble streamflow forecasting with diverse loss functions
- Open Access
- 31 March 2026
- Published by Elsevier in Journal of Environmental Management
- Vol. 404, 129508
Mapping soil pH baselines, trends, and time-to-critical crop risk using harmonized legacy datasets: A large-scale assessment
- 31 March 2026
- Published by Elsevier in Journal of Environmental Management
- Vol. 404, 129508
Open AccessMapping soil pH baselines, trends, and time-to-critical crop risk using harmonized legacy datasets: A large-scale assessment
- 24 February 2026
- Published by Wiley in European Journal of Soil Science
- Vol. 77 (2)
Soil Carbon Modeling at Crossroads: Building Reliable Methods for Policy and Practice
Soil carbon mapping (SCM) is rapidly becoming a cornerstone of soil science and environmental decision‐making, from precision agriculture to national carbon inventories. Yet SCM is at a crossroads: the methods that often promise high‐accuracy metrics can mask structural weaknesses that limit generalization and undermine policy relevance. Similar problems apply to larger Digital Soil Mapping (DSM). In this article, using soil organic carbon (SOC) as an illustrative example, we highlight three systematic sources of error that consistently inflate SCM performance: depth, bulk density, and spatial autocorrelation. Soil profile depth is often mishandled when profile increments are split between training and test sets, leading to inflated accuracy estimates. Bulk density (BD), essential for converting concentrations to stocks, is inconsistently applied and rarely accompanied by uncertainty estimates. SOC stocks at sampling locations are often derived using BD, and when BD is reintroduced as a predictor in machine learning models, it inflates reported accuracy and the model's predictive skill. Spatial autocorrelation further exaggerates accuracy when conventional random splits are used, while spatial blocking reveals much lower and more realistic predictive skill. Drawing on recent literature and our own analysis, we argue that SCM must adopt more rigorous practices, including profile‐level validation, spatially aware blocking, standardized reporting of assumptions, and alignment with policy‐relevant depth intervals. These steps will enhance comparability across studies and ensure that SCM outputs are credible for carbon accounting, climate mitigation, and land management purposes. The future of SCM and DSM depends on both new algorithms and methodological rigor and transparency.
- 24 February 2026
- Published by Wiley in European Journal of Soil Science
- Vol. 77 (2)
Soil Carbon Modeling at Crossroads: Building Reliable Methods for Policy and Practice
Soil carbon mapping (SCM) is rapidly becoming a cornerstone of soil science and environmental decision‐making, from precision agriculture to national carbon inventories. Yet SCM is at a crossroads: the methods that often promise high‐accuracy metrics can mask structural weaknesses that limit generalization and undermine policy relevance. Similar problems apply to larger Digital Soil Mapping (DSM). In this article, using soil organic carbon (SOC) as an illustrative example, we highlight three systematic sources of error that consistently inflate SCM performance: depth, bulk density, and spatial autocorrelation. Soil profile depth is often mishandled when profile increments are split between training and test sets, leading to inflated accuracy estimates. Bulk density (BD), essential for converting concentrations to stocks, is inconsistently applied and rarely accompanied by uncertainty estimates. SOC stocks at sampling locations are often derived using BD, and when BD is reintroduced as a predictor in machine learning models, it inflates reported accuracy and the model's predictive skill. Spatial autocorrelation further exaggerates accuracy when conventional random splits are used, while spatial blocking reveals much lower and more realistic predictive skill. Drawing on recent literature and our own analysis, we argue that SCM must adopt more rigorous practices, including profile‐level validation, spatially aware blocking, standardized reporting of assumptions, and alignment with policy‐relevant depth intervals. These steps will enhance comparability across studies and ensure that SCM outputs are credible for carbon accounting, climate mitigation, and land management purposes. The future of SCM and DSM depends on both new algorithms and methodological rigor and transparency.
- 21 March 2021
- Published by Springer Nature
- p. 43-50
Why Gamma Distribution of Seismic Inter-Event Times: A Theoretical Explanation
- 21 March 2021
- Published by Springer Nature
- p. 43-50
Why Gamma Distribution of Seismic Inter-Event Times: A Theoretical Explanation
- 21 March 2021
- Published by Springer Nature
- p. 37-41
Absence of Remotely Triggered Large Earthquakes: A Geometric Explanation
- 21 March 2021
- Published by Springer Nature
- p. 37-41
Absence of Remotely Triggered Large Earthquakes: A Geometric Explanation