Santosh Pant

Santosh Pant

Machine learning for Earth observation where ground truth is scarce: evaluation, calibrated uncertainty, and transfer across regions and sensors. B.S. Computer Science and Data Science, Knox College, expected June 2027.

EmailCVGoogle ScholarGitHubORCID

I am a senior at Knox College, finishing a B.S. in Computer Science and Data Science in June 2027, and I am applying to PhD programs for Fall 2027.

I work on machine learning for Earth observation where ground truth is scarce. My papers ask whether a benchmark number can be trusted, how to attach calibrated uncertainty to a prediction that someone will act on, and when a model trained in one region can be used in another. The work so far covers field boundaries, crop yield, radar mapping of monsoon rice, and drone thermal imagery of cities. Most of it is tested first in Nepal's Terai, where I grew up.

Two single-author papers were accepted in 2026, one at TMLR and one at ACM SIGSPATIAL. Most of this work was done independently. A third paper, a manuscript on RGB-to-thermal evaluation, comes from work with Prof. Geoffrey Siwo at the University of Michigan. I also build and ship software.

Upcoming: presenting [2] at ACM SIGSPATIAL 2026, Riverside, California, 3 to 6 November 2026, and [1] at the Midstates Undergraduate Research Symposium, Washington University in St. Louis, 6 to 7 November 2026.

The problem

A model that reads satellite images is only as trustworthy as the labels it was checked against. In the United States and Europe those labels exist. There are parcel registries, crop declarations, yield monitors on combines, and dense station networks. Global map products and geospatial foundation models are built and scored mostly against that kind of data. Most of the world has far less of it.

Map of Nepal and the Ganges Plain to its south. The plain is a faint band across Bihar and Uttar Pradesh, the 21 Terai districts of Nepal are shaded green, and Kanchanpur district in the far west is dark green.Indo-Gangetic PlainBiharMadhesh
Nepal's Terai districts on the northern edge of the Indo-Gangetic Plain. Kanchanpur, my home district, is marked.

Nepal's Terai is a clear case. Official paddy statistics come out by district, months after harvest. The monsoon hides most of the growing season under cloud. A model can score well on a global benchmark and nobody can say whether it works here, because there is so little to check it against. Most of my papers are about that gap. They ask what a reported number still means when the labels are proxies, the split leaks, or the baseline was weak.

The Terai is the northern edge of the Indo-Gangetic Plain. The same rice and wheat system continues across the border into Bihar and Uttar Pradesh, and on to Punjab and Bangladesh. It is farmed mostly by smallholders and it feeds several hundred million people. [2] also runs its pipeline on Bihar's 38 districts as an outside check. In the one year that could be tested, the model missed in the same direction in every district, and the paper reports that. I start in the Terai because I know what the ground looks like, and that helps most when the labels are wrong. The questions are not specific to farmland. [3] asks them of drone thermal imagery over cities.

Publications

Accepted

  1. Auditing GeoFM Evaluation for Field-Extent Segmentation: Label Proxies, Baselines, and When Frozen Features Match Fine-tuningSantosh Pant. Transactions on Machine Learning Research (TMLR), 2026. Single author. Accepted August 2026.Label proxies and baseline choice reorder geospatial foundation-model rankings for field-extent segmentation across six countries, and a frozen decoder is TOST-equivalent to full fine-tuning within 0.02 AUROC in 9 of 12 region-by-model cells while training about 110 times fewer parameters.OpenReview (reviews public)CodeArtifact (Zenodo)
  2. Calibrated District-Level Prediction Intervals for 2025 Paddy Yield Departure in Nepal's Madhesh ProvinceSantosh Pant. ACM SIGSPATIAL 2026, Applications track (4-page paper), Riverside, California, November 2026. Single author.Leak-controlled district-level intervals for Nepal's 2025 Madhesh drought from a five-model LightGBM ensemble with split and covariate-shift-weighted conformal calibration on Earth Engine features. At this calibration size the 2025 screen was inconclusive for all eight districts, and the paper says so.DOICode and provenance (Zenodo)

Manuscript

  1. Auditing Split Proximity and Radiometric Scale in RGB-to-Thermal EvaluationSantosh Pant, Adithya Vaithinathan Asokan, Haoyun Feng, Matthew R. Dennis, Keenan Gibbons, Verrah Akinyi Otiende, and Geoffrey Siwo. Manuscript, 2026. Available on request.Two training-free checks for RGB-to-thermal benchmarks. On CART, 97.8 to 100 percent of mixing-protocol test frames share a flight with a training frame, and per-frame normalization erases the physical temperature scale before scoring.

Research directions

  • Evaluation and audit methodology for remote sensing. Before trusting a benchmark number I ask whether it survives a different label source, a leakage-controlled split, or a stronger baseline. In [1], replacing field polygons with ESA WorldCover cropland as the label raises random-forest AUROC by up to 0.34, and a from-scratch U-Net beats a frozen Prithvi probe in five of six countries. The same question runs through [3].
  • Calibrated uncertainty for decision support. I build prediction intervals meant for district-level decisions and report when they are too wide to act on. [2] is the first case, for the 2025 paddy season in Nepal's Madhesh Province.
  • Transfer across regions and sensors. Labels are dense in some places and missing in others, so I test what a model trained where labels exist can still do where they do not. In my senior project, a model trained only on California crop-map labels is applied to NISAR radar images of Nepal's Terai and checked against an independent map in Nepal and field survey points in Pakistan (see Current work).

What I want to work on next

In a PhD I want to keep working on three questions. The first is evaluation that holds up when labels are sparse or second-hand, so a benchmark number means the same thing in Bihar as it does in Iowa. The second is uncertainty that the people using a map or a forecast can act on, which includes saying plainly when an interval is too wide. The third is transfer across regions and sensors, where a few hundred good labels in one place should be worth something in the next. I expect to keep testing these on South Asian agriculture. The methods are general, and I want to apply them to other Earth observation problems such as hazards, water, and land change. New sensors matter here too. NISAR's radar sees through monsoon cloud, and I use it in my senior project.

Current work

My Data Science senior project, supervised by Prof. Andrew Leahy (Mathematics, Knox College), uses NISAR L-band radar to map monsoon rice in Nepal's Terai. I rebuilt the Sentinel-1 rice rule that ICIMOD and Nepal's agriculture ministry use, for 21 Terai districts from 2021 to 2023. Against an independent map, it finds a median 32 percent of staple cropland and fails in the far west. I then trained a temporal attention model only on California crop-map labels and applied it to NISAR's 2026 images of Nepal. Inside the ministry's agricultural zone, it separates staple cropland from non-crop land better than the rule (AUROC 0.895 vs 0.746 in Bara, 0.878 vs 0.656 in Chitwan). On field points from a 2023 Asian Development Bank survey in Pakistan, it separates rice from non-crop land (AUROC 0.81 to 0.96) but not reliably from other summer crops, mostly sorghum (0.64 to 0.82).

From December 2026 to February 2027, I will be a Data Science Intern at the Smithsonian's National Zoo and Conservation Biology Institute, supervised by Dr. James Hassell (remote, part time). The project uses geospatial foundation models to classify watersheds and ecosystem types, starting in Kenya.

Other research

  • 2024 to 2026Richter Research Scholar, Knox College, four funded terms. Liquid-crystal image analysis, given as a talk at Knox Horizons, May 2026.
  • 2025Structured random graphs for HPC interconnects, with Prof. David Bunde, co-authored with Dipesh Adhikari. Presented at Knox Horizons, May 2025.

Software

I also build software. Two projects are below.

  • CropScan. Crop-disease diagnosis for small growers, built as a DINOv2 with LoRA and EfficientNetV2 ensemble that commits only when both models agree and otherwise asks for a better photo. Accuracy is 75.6 percent on real field photos and 99.7 percent on lab images. Three-person team. I built the ML pipeline and the backend. cropscan.techGitHub
  • Vigilant AI. Makes CCTV footage searchable with object detection, image-embedding search, and a vision-language model behind a chat interface. Built in 48 hours with one teammate. It won the Actian VectorAI DB sponsor track at HackIllinois 2026. GitHub

Honors

News

  • Nov 2026Presenting at ACM SIGSPATIAL 2026, Riverside, California.
  • Nov 2026Presenting at the Midstates Undergraduate Research Symposium, Washington University in St. Louis.
  • Oct 2026Accepted a Data Science internship at the Smithsonian's National Zoo and Conservation Biology Institute, December 2026 to February 2027.
  • Oct 2026Received an NSF-funded student travel award for ACM SIGSPATIAL 2026.
  • Aug 2026Paper accepted at ACM SIGSPATIAL 2026, Applications track.
  • Aug 2026Single-author paper accepted at TMLR.

Teaching and service

Teaching assistant for Programming Languages (fall 2026) and Data Structures and Algorithms (two terms), computer science tutor, Resident Assistant, and Computer Science Club lead (2025 to 2026) at Knox College.

Contact

Email is the fastest way to reach me: spant@knox.edu.

CVGoogle ScholarGitHubLinkedInORCID 0009-0000-6464-0777