TerraTrace: Temporal Signature Land Use Mapping System

Angela Busheska, Vikram Iyer, Bruno Silva, Peder A. Olsen, Ranveer Chandra, Vaishnavi Nattar Ranganathan

Abstract

TerraTrace is a novel land-use mapping system designed to distinguish between agricultural land and natural forests by analyzing "temporal signatures"—the unique patterns of vegetation greenness over a growing season. By integrating a massive longitudinal dataset for California with an LLM-powered analytics platform, TerraTrace provides a scalable, low-cost solution for tracking land-use changes and enforcing global anti-deforestation regulations like the EUDR.

Introduction

Global deforestation remains a critical driver of greenhouse gas emissions, with approximately 10 million hectares of forest cleared annually. Regulatory frameworks such as the European Union Deforestation Regulation (EUDR) now mandate that companies verify their products are not sourced from recently deforested land. However, current satellite monitoring tools face a significant "vision problem": from a static perspective, managed plantations and wild forests often appear identical, leading to frequent misclassifications.

Figure 1: (A) Visual similarity between evergreen forests and orchard trees leads to classification errors; (B) Feature comparison of TerraTrace against existing systems.

Methodology

TerraTrace shifts the analytical focus from static visual features to temporal behavior. The system utilizes the Normalized Difference Vegetation Index (NDVI) to track plant "greenness" over time. While natural forests exhibit stable or predictable seasonal cycles, agricultural crops display sharp signatures characterized by planting, rapid growth, and sudden harvest.

The system processes Sentinel-2 imagery, fitting data to polynomial curves to extract metrics such as peak growth timing and harvest-related decline rates. A unique feature of TerraTrace is its use of GPT-4 Turbo as a reasoning engine. The LLM acts as a domain expert, interpreting these NDVI curves and metrics to provide natural language summaries, bridging the gap between complex remote sensing data and actionable regulatory insights.

Data and Results

The researchers developed a novel California Longitudinal NDVI Dataset containing over 70 million data points from 2020–2023. Validation against the USDA's Crop Data Layer (CDL) confirmed the system's ability to identify unique signatures for various crops, including corn, citrus, and chickpeas.

Figure 2: TerraTrace successfully distinguishes between wheat fields and adjacent pine stands in Washington state.

In a critical test case in Washington state, TerraTrace successfully differentiated between a wheat farm and an adjacent pine stand—a scenario where traditional forest probability maps frequently fail. Furthermore, global tests on coffee farms in Vietnam and Honduras demonstrated that these biological "fingerprints" remain consistent across different geographic regions, suggesting high generalizability driven by plant biology rather than local geography.

Conclusion and Future Work

TerraTrace demonstrates that agricultural practices create distinct mathematical signatures that can reliably separate man-made farms from natural forests. While the current system is highly effective, future work will focus on expanding geographic coverage, increasing temporal resolution through harmonized data sources, and incorporating additional spectral indices like Land Surface Temperature to further refine vegetation fingerprints and monitor smaller, subsistence-style farms.

Summary updated

BibTeX
@inproceedings{busheska2025terratrace,
  author = {Angela Busheska and Vikram Iyer and Bruno Silva and Peder A. Olsen and Ranveer Chandra and Vaishnavi Nattar Ranganathan},
  title = {{TerraTrace: Temporal Signature Land Use Mapping System}},
  booktitle = {HotMobile},
  year = {2025},
  doi = {10.1145/3708468.3715679},
}
Figure