You are here

Harvard Forest Data Archive


Annual Maps of Forest Harvest Events in Maine from LANDSAT Imagery 1986-2019

Related Publications


  • hf437-01: Ensemble forest harvest maps and masks


  • Lead: Valerie Pasquarella, Jonathan Thompson
  • Investigators: Christopher Brown, John Kilbride, Luca Morreale
  • Contact: Information Manager
  • Start date: 1986
  • End date: 2019
  • Status: complete
  • Location: Maine
  • Latitude: +42.9778 to +47.4591 degrees
  • Longitude: -71.1114 to -66.9467 degrees
  • Elevation: 0 to 1606 meter
  • Datum: WGS84
  • Taxa:
  • Release date: 2023
  • Language: English
  • EML file: knb-lter-hfr.437.2
  • DOI: digital object identifier
  • EDI: data package
  • DataONE: data package
  • Related links:
  • Study type: historical
  • Research topic: conservation and management; ecological informatics and modelling; historical and retrospective studies
  • LTER core area: disturbance patterns
  • Keywords: clearcuts, conservation, disturbance, geographic information systems, history, land cover, land use, management, maps, remote sensing, timber harvest
  • Abstract:

    We used Landsat satellite imagery and forest inventory plot measurements to develop a time series of annual maps representing potential forest harvest events for the state of Maine in the Northeastern US for the years 1986 to 2019. We first generated a set of LandTrendr temporal segmentation results for three different spectral indices. Change results were filtered to remove events greater than two years in duration, then results were combined using a seven-parameter degenerate decision trees model that determined a set of thresholds on disturbance patch size, magnitude of spectral change, and change “votes” across indices. We found that we were able to detect harvest events that removed at least 30% of total basal area with a mean F1 score of 0.72 (σ = 0.02) with a mean false negative error rate (omission) of 0.32 (σ = 0.02) and mean false positive error rate (commission) of 0.23 (σ = 0.03), and these scores further improve when maps are masked to remove human land use (built and agriculture) and water based on National Land Cover Dataset and JRC Global Surface Water classifications (mean F1 = 0.73, σ = 0.02). Comparisons with an out-of-sample reference dataset and an existing national forest disturbance dataset indicate our forest harvest maps are a locally accurate source of information for characterizing spatial and temporal variability in long-term harvest patterns across the industrial forests of northern Maine. Here, we provide annual ensemble-based maps of potential harvest events; cross-validated results, which give an indication of detection agreement across subsets of our forest inventory reference datasets; and ancillary datasets that can be used to mask false detections in urban and agricultural land uses and water.

  • Methods:

    Temporal segmentation and ensemble change detection

    We used the Google Earth Engine implementation of LandTrendr temporal segmentation approach to generate inputs for our harvest event detection ensemble. We applied LandTrendr to annual medoid composites of all high-quality Landsat 5, 7 and 8 Collection 1 Surface Reflectance observations acquired between June 20 and September 20 (Northern Hemisphere growing season) for the years 1985–2020. Initial segmentation results were generated separately for three SWIR-based indices, (1) the Normalized Burn Ratio (NBR), (2) the Normalized Difference Moisture Index (NDMI), and (3) Tasseled Cap Wetness (TCW). LandTrendr outputs include a series of segments, which correspond to relatively stable periods, and vertices, which were identified as inflection points along a spectral trajectory and are indicative of potential changes in surface conditions. We considered all loss vertices, i.e., those with spectral changes in the direction of decreased vegetation cover, as potential disturbance events. To differentiate harvests from longer-duration disturbances such as those related to drought or forest insect damage, we removed vertices associated with segments greater than two years in duration, leaving only short-term (less than 2 year) events that are more likely associated with harvesting.

    We used all available FIA field plot measurements collected in the state of Maine between 1999 and 2019 to train and cross-validate harvest detection ensembles and we had access to true plot locations through a memorandum of understanding between the USFS and Harvard University (MOU #09MU11242305123). Our Maine FIA dataset consisted of 13,299 measurements (i.e., unique space-time coordinates) recorded for 3,265 plots (i.e., unique spatial locations), and of these, we analyzed the 3,220 FIA plots that had been remeasured at least once and our final dataset included 10,034 pairs of sequential FIA measurements, of which 1,711 recorded basal area removal (harvest).

    To integrate the Landsat-based and FIA datasets, we queried the LandTrendr results for all years between the first and second FIA measurement years to determine if a potential harvest event was detected between measurements. If multiple events (vertices) were detected in a given measurement period, we recorded information for the event with the greatest magnitude of change. We excluded the first year of each measurement pair to prevent double-counting endpoint years for plots with two or more remeasurements. The resulting dataset included a record for each FIA remeasurement with plot information from each measurement pair as well as the LandTrendr features for each of the three spectral indices we considered.

    We used a degenerate decision trees approach to combine LandTrendr results for different spectral indices into a final time series of forest harvest event maps. Degenerate trees are a subclass of binary trees where each decision node has only a single parent node. We provide an example of our DDT implementation at, including a Python notebook with example functions for running a sweep over series of thresholds and determining optimal thresholds for each feature as well as an Earth Engine script for applying thresholds to LandTrendr results.


    We used our FIA remeasurement dataset to perform a three-fold cross validation replicated ten times for a total of 30 folds. We also performed a second validation using TimeSync interpretations collected as part of another effort to model change processes in the northeastern US and Canada. This dataset comprised 3,436 TimeSync interpretations representing 1,294 unique spatial locations within our Maine study area including 634 events labeled harvest. Finally, we compared our ensemble results with annual LCMS disturbance maps (v2020-5) from 1985-2020 generated as part of the USFS Landscape Change Monitoring System (LCMS) project. These national products do not differentiate among forest harvest and other types of “fast loss” but represent an existing national-scale alternative to the the dataset presented here. We consistently found our datasets outperformed LCMS products for forest harvest detection in FIA cross-validation (mean F1 = 0.72 compared with F1 = 0.60) and for the TimeSync reference dataset (F1 = 0.64 compared with F1 = 0.57 for harvest detection). These assessments indicate our dataset represents a best-available source of information on Maine harvest regimes over the last three decades.

    Annual harvest event maps

    Ensemble harvest detection results are provided as annual binary change maps (v6-0_ensemble_ba_perc). For each harvest event map, 0 indicates no harvest detected and 1 indicates potential harvest detected.

    Cross-validated annual harvest event maps

    We also include cross-validated map results (v6-0_ensemble_ba_perc_cv), which are generated by parameterizing ensembles on a subset of FIA measurements and provide a measure of uncertainty in terms of percent agreement across the ensemble of ensembles (n=30).


    Given the focus on forest harvest, harvest event maps can be further filtered (masked) to remove events detected in non-forested land use types, specifically urban (built), agriculture (crops), and water, and masks derived from the 2016 National Land Cover Datasets (NLCD) and JRC Global Surface Water annual layers are included for use in post-processing ensemble harvest detection products.

    masks_ME.tif: Water and non-vegetated (built, agriculture) masks. Band 1: gsw_mask, Band 2: nlcd_mask, Band 3: combined_mask

    nlcd2016_ME.tif: NLCD 2016 for Maine. Can used to build alternate masks.

  • Organization: Harvard Forest. 324 North Main Street, Petersham, MA 01366, USA. Phone (978) 724-3302. Fax (978) 724-3595.

  • Project: The Harvard Forest Long-Term Ecological Research (LTER) program examines ecological dynamics in the New England region resulting from natural disturbances, environmental change, and human impacts. (ROR).

  • Funding: National Science Foundation LTER grants: DEB-8811764, DEB-9411975, DEB-0080592, DEB-0620443, DEB-1237491, DEB-1832210.

  • Use: This dataset is released to the public under Creative Commons CC0 1.0 (No Rights Reserved). Please keep the dataset creators informed of any plans to use the dataset. Consultation with the original investigators is strongly encouraged. Publications and data products that make use of the dataset should include proper acknowledgement.

  • License: Creative Commons Zero v1.0 Universal (CC0-1.0)

  • Citation: Pasquarella V, Thompson J. 2023. Annual Maps of Forest Harvest Events in Maine from LANDSAT Imagery 1986-2019. Harvard Forest Data Archive: HF437 (v.2). Environmental Data Initiative:

Detailed Metadata

hf437-01: Ensemble forest harvest maps and masks

  • Compression: zip
  • Format: tiff
  • Type: image