Monitoring Site Rehabilitation Using Landsat Time Series and Matched Reference Sites

The rehabilitation of cleared sites, such as well pads, is a core component of environmental restoration programs. Monitoring vegetation recovery over time is essential for assessing restoration effectiveness, and identifying when additional intervention may be needed. Remote sensing provides significant advantages over field-based methods by enabling repeatable, cost-effective observations at scale. In this project, we focused on the analysis of open-source satellite products. While commercial VHR imagery can add significant value to targeted assessments, satellite programs such as Landsat and Sentinel remain the foundation for consistent, long-term, and cost-effective monitoring of rehabilitation outcomes.

In this study, we leveraged Landsat products, which are jointly managed by NASA and USGS [1, 2]. Landsat’s global coverage, consistent spectral bands, and long temporal depth make it particularly suited for long-term environmental change detection. Our analysis begins with Landsat 5, which provides 30m spatial resolution, across 7 spectral bands, and a 16-day revisit cycle, from 1984. Critical for rehabilitation assessments, this allowed us to remote sense the pre-clearance state of the target sites.

Several recent studies support the value of Landsat time series, reference-based assessment, and the application of statistical approaches to ecological monitoring. For example, Monroe et al. [3] developed a dynamic reference approach using NDVI to monitor sagebrush recovery, at former oil and gas well pads. Wang et al. [4] tracked vegetation disturbance and restoration in open-pit coal mines, using Landsat products and the Lantrendr temporal segmentation algorithm in Google Earth Engine [5].

We can divide our methodology into four phases, as follows:

1. Site Pairing for Reference Selection

To control for spatial and environmental variability, each cleared rehabilitation site was paired with three nearby reference sites. Firstly, candidate reference pixels were selected, based on spatial proximity to the rehab site, pre-clearance land cover class, and temporal stability. The latter was calculated using standard deviations of fractional cover components from DEA [6], the goal being to avoid reference sites that were subject to anthropogenic disturbance.

Secondly, candidate reference sites were filtered and ranked, using Euclidean distance, across a range of normalized environmental and spectral features, including topographic relief, various soil metrics from TERN [7], and pre-clearance fractional cover Fig.1.

Fig.1 Top: Radar plots of the rehab site vs. its matched paired sites based on pre-clearance fractional cover. Bottom: Bar plot of the rehab site vs. its matched paired sites based on soil covariates.

2. NDVI Time Series Extraction

For both the rehab site and its reference sites, median annual Normalized Difference Vegetation Index (NDVI) values were extracted from Landsat surface reflectance products, from 1984 to the present, using GEE [8]. This extended time series encapsulated the pre-clearance state at our rehab site. To reduce cloud contamination and phenological noise (natural variability in vegetation signals caused by seasonal growth and senescence cycles), the analysis was restricted to the southern hemisphere growing season: Nov to March of the following year Fig.2.

Fig.2 Example NDVI time series for a well pad site. The site extent is outlined in red. The site was cleared in 2010. Note the visible NDVI reduction and apparent recovery, post-clearance.

3. Recovery Threshold and Logistic Modelling

Using pre-clearance NDVI values from the rehab site, and its matched reference sites, a recovery threshold was calculated, defined here as 90% of the pre-disturbance NDVI. Post-disturbance NDVI values at the rehab site were fitted to a regression model. The year in which the modelled NDVI curve exceeded the recovery threshold was recorded as the predicted recovery year. This analysis was performed on both raw and smoothed NDVI values Fig.3.

Fig.3 Example NDVI time series with rehab trajectories, based on log regression post-clearance (raw and smoothed data shown). The clearance date is indicated in red, the predicted recovery date is shown in green.

4. LSTM-Based Prediction of Rehabilitation Trajectories

To enhance rehab trajectory predictions, a Long Short-Term Memory (LSTM) neural network was trained on a dataset of 144 rehabilitated sites. The LSTM architecture is capable of capturing nonlinear temporal dependencies, using both dynamic (e.g., weather, pre-clearance NDVI) and static (e.g., soil, relief) site features [9]. The LSTM was trained to predict raw post-clearance NDVI trajectories, over 5-fold cross-validation, and out-of-fold predictions were saved for evaluation. Preliminary results, over our small dataset, suggest the potential for learning typical vegetation recovery trajectories across sites, and weather scenarios. I.e. generalising rehab trajectory experience (Fig.4).

Fig.4 Example post-clearance NDVI time series predictions from LSTM (out of fold).

In conclusion, this method used the Landsat archive to provide an efficient, scalable, and ecologically significant assessment of rehabilitation progress at cleared well pads. The exceptional temporal depth of the Landsat archive allowed us to directly view the pre-clearance state of target sites, and thereby establish reasonable matched pairs, and post-clearance rehab targets. The integration of LSTM modelling may enhance the robustness of recovery estimates. For a more complete picture of site rehabilitation, NDVI-based time series approaches should be combined with other remote sensing modalities, including those using VHR satellite products.

References

[1] USGS (2021). "Landsat Satellites": https://landsat.gsfc.nasa.gov/satellites/

[2] Wulder, M. A., et al. (2019). "Current status of Landsat program, science, and applications." Remote Sensing of Environment, 225, 127–147. https://doi.org/10.1016/j.rse.2019.02.015

[3] Monroe, A. P., et al. (2022). "Assessing vegetation recovery from energy development using a dynamic reference approach." Ecological Indicators, 141, 109135.

[4] Wang, Y., et al. (2021). "Tracking the vegetation change trajectory over large-surface coal mines in the Jungar coalfield using Landsat time-series data." Remote Sensing, 13(18), 3625.

[5] Pasquarella, V. J., et al. (2022). "Demystifying LandTrendr and CCDC temporal segmentation." Remote Sensing of Environment, 276, 113124.

[6] Digital Earth Australia products. https://explorer.dea.ga.gov.au/products

[7] Terrestrial Ecosystem Research Network, TERN Soil and Landscape Grid of Australia. Dataset accessed via TERN Data Discovery Portal. Available at: https://portal.tern.org.au

[8] Google Earth Engine Team (2015) Google Earth Engine: A planetary-scale geospatial analysis platform. Available at: https://earthengine.google.com

[9] Hochreiter, S. and Schmidhuber, J. (1997) ‘Long short-term memory’, Neural Computation, 9(8), pp. 1735–1780. doi:10.1162/neco.1997.9.8.1735.