Projects | Xiaojie Gao

Mapping Forest Communities at 10-m Resolution Using Deep Learning

Wed, 10 Sep 2025 00:00:00 +0000

Abstract

Collaborating with Google DeepMind and Earth Engine teams, we map forest communities at 10-m resolution across Northeastern and Midwestern United States using a novel remote sensing embedding fields dataset generated by deep learning and in-situ forest inventory measurements. We provide community-constrained species maps that delineate the spatial distributions of more than 50 tree species, along with maps of plot-level aboveground carbon, basal area, and stem density. These products offer a foundation for applications in landscape simulation, biodiversity conservation, and socio-economic policymaking.

To provide an early access, please see our preprint: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5936862

Multiscale Analysis of Land Surface Phenology

Mon, 01 Apr 2024 00:00:00 +0000

The paper is published in Remote Sensing and Environment !

Abstract

Land surface phenology (LSP) metrics derived from remote sensing are widely used to monitor vegetation phenology over large areas and to characterize how the growing seasons of terrestrial ecosystems are responding to climate change. Until recently, however, most LSP studies relied on coarse spatial resolution sensors, which makes assigning direct linkages between LSP metrics and ecological processes and properties challenging due to scale mismatches and because substantial variation in phenology and ecological properties are often present at sub-pixel scale in coarse resolution LSP metrics. In this study, we leverage publicly available LSP data products with three orders of magnitude difference in spatial resolution derived from Moderate Resolution Imaging Spectroradiometer (MODIS, 500 m), Landsat and Sentinel-2 (HLS, 30 m), and PlanetScope (3 m) imagery to examine and characterize the nature, magnitude, and sources of the agreement and disagreement in LSP metrics across spatial scales. Our results provide three key conclusions: (1) LSP metrics from three sensors showed consistently high cross-scalar agreement across sites (r2 = 0.70 – 0.97), suggesting that they all effectively capture geographic variation in LSP; (2) within-site cross-scalar agreement between LSP metrics was systematically lower relative to agreement across sites, but mean absolute differences were consistent across and within sites (generally <14 days for day of year-based metrics, with a few exceptions); and (3) local-scale composition and heterogeneity in land cover is a key factor that controls cross-scalar agreement in LSP metrics. In particular, we found that site-level heterogeneity in land cover (measured via entropy) and the proportion of evergreen versus deciduous land cover types explain up to half of site-to-site variance in local-scale cross-scalar agreement in LSP metrics. Results from this study support the internal consistency and quality of the three LSP data products examined, and more generally, provide guidance regarding the choice of spatial resolution for different applications and land cover conditions, and yield new insights related to how LSP observations scale across different sensors and spatial resolutions.

PnET-R

Sun, 01 Oct 2023 00:00:00 +0000

Ecosystem models offer a rigorous way to formalize scientific theories and are critical to evaluating complex interactions among ecological processes. In addition to simulation and prediction, ecosystem models are a valuable tool for testing hypotheses about mechanisms and empirical findings because they reveal critical internal processes that are difficult to observe. However, many ecosystem models are difficult to manage by ecologists for scientific exploration due to complex model structures, lack of consistent documentation, and low-level programming implementation, which facilitates computing but reduces accessibility. Here, we present the pnetr R package to provide an easy-to-manage ecosystem modeling framework and detailed documentation in both model structure and programming. The framework implements a family of widely used PhotosyNthesis and EvapoTranspiration (PnET) ecosystem models, which are relatively parsimonious but capture essential biogeochemical cycles of water, carbon, and nutrients. We chose the R programming language since it is familiar to ecologists and has abundant statistical modeling resources. We showcase examples of model simulations and test the effects of phenology on carbon sequestration and wood production using data measured by the Environmental Measurement Station (EMS) eddy-covariance flux tower at Harvard Forest, MA. We hope pnetr can facilitate further development of ecological theory and increase the accessibility of ecosystem modeling and ecological forecasting.

The paper is published in Methods in Ecology & Evolution. See details here

The pnetr package is open source: https://github.com/hf-thompson-lab/pnetr

DISES: Co-produced modeling of socio-environmental dynamics of financialized forestlands and alternative future scenarios

Tue, 01 Aug 2023 00:00:00 +0000

Abstract

Financial firms have taken ownership of tens of millions of hectares of commercial forestlands in the United States in recent decades. However, the governance, community access, and ecological implications of this change remain largely unknown. This project will advance scholarship on timberlands in the United States as integrated socio-environmental systems, understand what drives decision-making for finacialized lands, the social and environmental implications of these landscape changes, and inform policy deliberations regarding their future ownership and management.

This project will develop spatial and causal models to represent the socio-environmental feedbacks and consequences of three decades of forest financialization, then use these models to co-produce future landscape scenarios for informing policy and governance processes. Specifically:

Analyze the socio-environmental implications arising from the industrial timberland financialization process in terms of landscape pattern changes, harvest patterns, carbon storage, and access to forest benefits by tribal and non-tribal forest users.
Identify the actors, processes, and variables that act as leverage points in driving changes in patterns of carbon storage, forest conversion, and community benefits within timberlands as socio-environmental systems.
Develop and simulate future governance and management scenarios for these lands that evaluate their carbon, ecological, and social implications and tradeoffs through a collaborative co-production process with key stakeholders.

Our collaborators include scientists and researchers in the University of Maine, the University of Georgia, and Google LLC.

Does chilling explain the divergent response of spring phenology to urban heat islands?

Wed, 18 Aug 2021 00:00:00 +0000

This NASA-funded project aims to test the hypothesize that the divergent spring phenology trends in urban areas can be explained by the interaction between UHI-induced seasonal temperature changes and variable plant chilling requirements—the need of plants to be exposed to sufficiently low temperatures to release dormancy in spring.

Urbanization is known to have direct impacts on plant phenology. Understanding these effects is important to biodiversity dynamics, ecosystem structure, carbon cycles, and human health. Temperature increases from the Urban Heat Island (UHI) effect are thought to be the main driver of plant phenological changes around cities. However, trends in plants’ start of growing season (SOS) dates around urban areas, compared to the surrounding countryside, have diverged across the globe: some advance, and some delay. Divergent SOS trends have been observed in field measurements as well as satellite remotely sensed terrestrial vegetation seasonality—land surface phenology (LSP). However, the reasons for this phenomenon remain unclear. We hypothesize that divergent SOS trends can be explained by the interaction between UHI-induced seasonal temperature changes and variable plant chilling requirements—the need of plants to be exposed to sufficiently low temperatures to release dormancy in spring. The proposed project is designed to evaluate this hypothesis by accomplishing two main objectives:

• Objective 1: Map long-term medium spatial resolution LSP for large cities.

More than 50 large cities in North America that are expected to have strong UHI effects will be selected to capture variability in biome type and prevailing climate.
LSP with pixel-wise uncertainty at 30 m spatial resolution from 1984 to present will be retrieved for each selected city using a recently developed Bayesian model.

• Objective 2: Test chilling hypothesis by analyzing spring phenology models

A suite of models describing spring emergence as a function of daily temperature exposures will be fit using the generated LSP data.
Analysis of model fit and parameters will indicate the importance of UHI-altered chilling/warming regimes.

We expect this work will make contributions to the remote sensing, phenology, and global change science communities by providing:

An improved understanding of how temperature controls plant phenology in urban areas;
A novel approach to retrieve medium spatial resolution LSP with pixel-wise uncertainty that can be generally applied to urban or natural regions,
A long-term 30 m spatial resolution LSP dataset for large cities in North America.