Demystifying LandTrendr and CCDC temporal segmentation

Improved access to remotely sensed imagery and time series algorithms in combination with increased availability of cloud computing resources and platforms such as Google Earth Engine have significantly expanded the community of users processing and analyzing time series of satellite observations. Though individual time series analysis methods and their applications tend to be well-documented, comparisons of different approaches are beneficial to new users faced with the choice of different algorithms and parameterizations. We review two temporal segmentation approaches that have become increasingly prevalent in land cover mapping and monitoring: LandTrendr (Landsat-based detection of Trends in Disturbance and Recovery) and CCDC (Continuous Change Detection and Classification). We examine differences in the way these approaches use the temporal and spectral domains and compare model specifications and outputs. This review highlights previous work and applications, current limitations, ongoing challenges, and opportunities for future integration and comparison of methods and map products, and is expected to benefit both user and developer communities.


Introduction
With unprecedented access to satellite imagery and cloud-based processing capabilities, more users are working with multi-temporal Earth observation datasets than ever before Wulder et al., 2012). Following the opening of the Landsat archive in 2008, time series-based approaches for land cover mapping and change detection have become increasingly common and remain an active area of development, both in terms of methods and theory (Banskota et al., 2014;Chaves et al., 2020;Gómez et al., 2016;Hemati et al., 2021;Woodcock et al., 2020). Though more granular taxonomies for characterizing change detection approaches have been proposed (e.g., Zhu, 2017), we consider the broad class of temporal segmentation algorithms to include methods that group observations into larger representative periods of time or "segments" such that a single model is applied within a segment and different models are applied across segments. These piecewise fitting approaches can be viewed as a temporal equivalent to spatial segmentation algorithms, which use spectral properties to group contiguous pixels into shapes representing larger landscape objects or patches (e.g., Hossain and Chen, 2019). Characterization of temporal segments enables analysis of both short-and long-term trends and patterns in landscape properties and surface conditions. However, the types of change processes and events that can be identified are intrinsically linked to the algorithm used to perform temporal segmentation .
Two temporal segmentation approaches that are becoming more prevalent in natural resource monitoring applications are the Land-Trendr (Landsat-based detection of Trends in Disturbance and Recovery) and CCDC (Continuous Change Detection and Classification) algorithms (Zhu, 2017). Though the use of these approaches was once limited to a small subset of users with sufficient data storage and processing capabilities, both LandTrendr (Kennedy et al., 2010) and CCDC (Zhu et al., 2012;Zhu and Woodcock, 2014) have since been implemented in Google Earth Engine (Arévalo et al., 2020a;Kennedy et al., 2018b). Implementation in Earth Engine's cloud computing environment has enabled a global community of users to run these and other temporal segmentation methods, such as Exponentially Weighted Moving Average Change Detection (EWMACD; Brooks et al., 2014), Vegetation Change Tracker (VCT; Huang et al., 2010) and Vegetation Regeneration and Disturbance Estimates through Time (VeRDET; Hughes et al., 2017). Furthermore, both LandTrendr and CCDC are being used operationally for large-area mapping and monitoring efforts (e.g., Brown et al., 2019;Kennedy et al., 2018a;Lister et al., 2020;Tarrio et al., 2019;Xian et al., 2021;Zhu et al., 2016). Therefore, we anticipate continued increases in the use of these algorithms as well as their derivative products.
With a growing user community, it is increasingly important to build a comprehensive understanding of the assumptions, advantages, and limitations of individual algorithms as well as differences across approaches (Zhu, 2017). In the case of LandTrendr and CCDC, both algorithms partition time series of observations into sequences of modeled temporal segments, both can be used to produce annual land cover and land cover change map products, and both can generate smoothed or "synthetic" imagery. However, these algorithms fundamentally differ in their segmentation approaches, interpretation of the resulting model fits, and output datasets. Quantitative assessments using independent reference data for specific change detection applications have been addressed in other studies such as Cohen et al. (2017), who present a more comprehensive comparison of forest disturbance maps generated using different temporal segmentation approaches. In this paper, we use cases from previously published studies as well as illustrative, singlepixel examples to review key decisions in working with the Land-Trendr and CCDC algorithms. This focused, in-depth perspective is intended to aid users in interpreting results and products and serve as an inroad for further comparisons with other temporal segmentation approaches.

Overview of algorithms
The LandTrendr and CCDC algorithms are both well-documented in previous publications (see Kennedy et al., 2012Kennedy et al., , 2010Saxena et al., 2018;Zhu and Woodcock, 2014), and code is available in open-source repositories (e.g., Table 1). Our objective here is not to revisit the mechanics of these mature and widely cited temporal segmentation algorithms, but to review fundamental design choices and sensitivities to user-specified inputs. Specifically, we consider how LandTrendr and CCDC use the temporal domain for segmentation (Section 2.1), the influence of temporal and spectral inputs on model fitting (Sections 2.2 and 2.3), and additional considerations in selecting parameter inputs (Section 2.4). We also summarize key model outputs and derived data products (Section 2.5) and related visualization tools (Section 2.6).

Use of the temporal domain
Landscape change processes occur at many time scales and may include intra-annual (seasonal) changes in phenology and condition, as well as inter-annual processes such as abrupt shifts in land cover, longterm trends, and more complex disturbance-recovery trajectories (Coppin et al., 2004;Pasquarella et al., 2016;Vogelmann et al., 2016). The frequency of observations used to detect and model temporal dynamics directly influences the types of processes that can be characterized using time series of remotely sensed observations . Though temporal segmentation methods can be generalized to any time series of continuous values, LandTrendr and CCDC were both originally developed to take advantage of the unparalleled temporal coverage of the Landsat record and both algorithms typically use full Landsat collections as inputs. However, there are notable differences in how they reduce and analyze an observation record for a given pixel.
The LandTrendr approach is typically applied to time series data where one observation or value is selected to represent each year. Standard use of medoid composites is intended to minimize the influence of vegetation phenology, effectively controlling for seasonal variability prior to modeling temporal trajectories, and spectral values from different Landsat sensors are harmonized using transformations proposed by Roy et al. (2016). LandTrendr segmentation interprets input time series as continuous piecewise trajectories, which are typically described in terms of "segments" and "vertices" (Fig. 1). Because LandTrendr segmentation is performed on time series with an annual time step, change processes are associated with a year of detection rather than a specific date (month/day) of change.
In contrast, CCDC segmentation is typically performed using a time series of all clear observations. This approach attempts to detect change at the native frequency of the input time series such that changes can be associated with a particular acquisition date rather than a composite year. To explicitly account for intra-annual cyclic patterns in surface reflectance signals resulting from changes in illumination conditions, atmospheric conditions, and vegetation phenology, CCDC uses linear harmonic models that include sine and cosine terms. These harmonic terms eliminate the need to pre-specify a compositing period and provide a larger set of coefficients/features that can be used for segmentbased classification (e.g., Pasquarella et al., 2018;Zhu et al., 2016;Zhu and Woodcock, 2014). Unlike LandTrendr fitting, where vertices represent shared segment start/end dates, CCDC fits input time series using a discontinuous piecewise model consisting of independent segments and "breaks," where breaks correspond to discrete change events represented as jumps or offsets in the temporal trajectory ( Fig. 1). In some cases, harmonic model fitting may be unstable because of the small numbers of observations used in fitting, leading to extended gap periods between CCDC model segments (Fig. 2). Forward-or back-filling of CCDC parameters from adjacent segments or use of a "disturbed" category in segment-based classification can be applied to reconcile these gap periods. In comparison, LandTrendr will always produce a continuous trajectory.
Because both algorithms assume a simple linear relationship between time and reflectance (or other modeled value), they are wellsuited for identifying periods of stability or relatively stable rates of change, and segments can be characterized in terms of direction of change (stable, loss or gain) and sorted in terms of change magnitude and duration. However, nonlinear dynamics, that is, periods with increasing, decreasing, or otherwise nonstationary rates of change, are more challenging to approximate using linear basis functions (Fig. 2). Thus, in addition to discrete events, such as abrupt disturbances and changes in land cover, breaks and vertices may more generally represent approximate points of inflection in the rate of the fitted trajectory.

Temporal considerations
Ability to fit temporal trajectories and identify disturbances and Table 1 Examples of LandTrendr and CCDC implementations.
other change processes is largely determined by the number of clear observations available for analysis at both scene and pixel levels . Long-term satellite observation records like the Landsat archive vary both spatially and temporally because of clouds, shadows, snow/ice masking, scan line gaps, orbital path overlap, and image processing levels and archiving (Egorov et al., 2019;Wulder et al., 2016). Thus, when performing temporal segmentation, users should always consider ecosystems and/or processes of interest and site-specific patterns in observation frequency, including regularly occurring seasonal gaps (like those that occur in high latitudes because of low solar zenith angles) as well as periods of missing data corresponding to archival gaps, and adjust their choice of segmentation algorithm and algorithm-specific parameters accordingly.
Because LandTrendr segmentation relies on a single observation or value per year, appropriate selection of compositing period start and end dates is critical, and "best available pixel" composites for regions of interest at high latitudes are more likely to encounter issues related to phenology, whereas composites for locations closer to the equator are more likely to encounter issues related to persistent cloud cover (Banskota et al., 2014). Examples of compositing periods for previous studies using LandTrendr are provided in Table 2. Studies focused on Northern Hemisphere forest dynamics typically define the compositing period to approximate the growing season, with start dates in June or July and end dates in August or September (e.g., Frazier et al., 2015;Griffiths et al., 2012;Kennedy et al., 2010;Pflugmacher et al., 2012;Senf and Seidl, 2021). However, growing season start and end dates must be adjusted when considering vegetation dynamics in the Southern Hemisphere (e.g., Mugiraneza et al., 2020;Nguyen et al., 2018). The compositing period may also be set to characterize finer-grained local seasonality (e.g.,

Table 2
Examples of studies using LandTrendr temporal segmentation, including process of interest, study area, temporal compositing period and years analyzed, and spectral or other time series inputs used for segmentation.  , de Jong et al., 2021;Rathnayake et al., 2020;Xiao et al., 2020;Ye et al., 2021). Both the number of fitted LandTrendr segments and timing of vertices can vary depending on how the compositing period is specified (Fig. 3). Therefore, users are encouraged to experiment with composite start and end dates for example sites prior to performing segmentation at landscape scales. In some cases, users might also consider modifying the compositing method. For example, maximum spectral values might be more useful than medoid composites in agricultural settings or other ecosystems where seasonal cycles are less regular.
Although CCDC attempts to fit all available observations for a given pixel (Fig. 3), users must still consider how spatial and temporal variability in the frequency of observations may affect segmentation results. Examples of time periods used in previous studies employing the CCDC segmentation approach are shown in Table 3. Because CCDC identifies break points based on the consecutive number of observations that deviate from the predicted trajectory of the current segment, when there are long periods of missing data (e.g., full years with no observations), there will be a greater tendency to find breaks after these periods, resulting in an artificially inflated number of changes associated with post-gap acquisition dates. In these cases, users may choose to remove . Both algorithms are relatively consistent in identifying the timing of the initial change in cover for both sites, with the exception of the October-December period for the Egypt example, where no breaks are identified. Note how the number and placement of subsequent LandTrendr vertices varies depending on the compositing period while only a single result is obtained for CCDC since the CCDC approach uses all available observations and does not require users to specify a compositing period. years with sparse acquisition records from analysis (e.g., Arévalo et al., 2020b). Similarly, increases in the frequency of observations results in a shorter period of time needed to flag a change. For example, Brown et al., (2019) found that when running CCDC on Landsat Analysis Ready Data (Dwyer et al., 2018;Egorov et al., 2019), change detection rates increased in orbital path overlap (sidelap) zones and during periods when multiple satellites are in operation, such as from April 1999 to November 2011 when Landsats 5 and 7 were both in operation. These effects were mitigated by dynamically adjusting the number of observations needed to detect a change; however, this custom modification has not been incorporated into all versions of CCDC and is not currently available in the Earth Engine implementation. Thus, Earth Engine users working with merged Landsat collections across larger study areas spanning multiple scenes should consider spatial and temporal inconsistencies and may want to apply additional filters on path overlaps and row endlaps or run separate models for individual scene-based image stacks. Changes in observation density and irregularity also affect the use of harmonic regression models for time series reconstruction (i.e., generating synthetic images), and a data density of at least seven observations per year is recommended for Landsat time series analysis (Zhang et al., 2021).
Another important set of considerations for temporal segmentation are the time series start and end dates and length of time series being segmented. LandTrendr is considered an "offline" approach in that the full series of composited values is used to determine the optimal set of segments and vertices (Zhu, 2017). The user must specify the maximum number of segments to use in model fitting, and changing the time series start and end dates can affect the trajectory fitting and timing of vertices (Fig. 4). Changing the number of segments used for fitting can also result in differences in fitted trajectories, and it is generally advisable to increase the number of allowable segments in studies of frequently disturbed landscapes. Number of segments should also generally be increased relative to the length of time series records to account for increasing likelihood of multiple disturbance events occurring over longer time periods.
Unlike LandTrendr, CCDC employs an "online" approach to segmentation such that observations are considered in an iterative, forward-looking manner when determining the placement of breaks (Zhu, 2017). As a result, CCDC models are sensitive to values at the beginning of fitting periods, and stability tests are included in the fitting procedure to encourage more stable model initialization . Changing the start year of input time series can significantly influence the timing of breaks as well as the fit of harmonic segments, especially if monitoring begins during a relatively unstable period (e.g., Fig. 4). CCDC segmentation results may also change as new dates of imagery are added to update time series and results, with longer CCDC segments based on greater numbers of observations and strong temporal signatures tending to be more robust to the addition of new observations.

Table 3
Examples of studies using CCDC and CCDC-like temporal segmentation, including process of interest, study area, use of the temporal domain, and spectral inputs used for break detection.

Use of the spectral domain
Another fundamental consideration when performing temporal segmentation is the choice of spectral inputs. When applied to Landsat time series, typical inputs include surface reflectance and brightness temperature measurements (i.e., Landsat bands) and multi-spectral indices and transforms, though time series from other sensors and of other continuous values such as classification probabilities or climatic variables can also be used.
LandTrendr was designed as a univariate model, where only one band, index, or feature is used in segmentation. It is important that users select an input that best captures the change process of interest (see Table 2 for examples). The original LandTrendr implementation (Kennedy et al., 2010) was tested using the Normalized Burn Ratio (NBR),   Fig. 1 (45.4189 • , − 69.8050 • ). CCDC models were fit in a multi-spectral context, resulting in break information that is shared across input features/bands, as illustrated by the consistent number of segments and locations of breaks in CCDC results. In contrast, the LandTrendr algorithm is univariate, with segmentation performed on a single band, index, or other continuous value, resulting in different numbers of segments and placement of vertices, depending on the input. Tasseled Cap Wetness (TCW), and the Normalized Difference Vegetation Index (NDVI), and these indices remain common choices (Table 2). Comparisons by Cohen et al. (2020) and Cohen et al. (2018) suggest that SWIR-based features tend to exhibit the strongest performance for forest disturbance detection in a univariate context, with TCW having the lowest overall error rates. Given the sensitivity of NDVI to vegetation cover, NDVI time series have also been employed in studies using LandTrendr to characterize mining disturbances (Xiao et al., 2020;Yang et al., 2018), cropland conversion  and more general land cover change (Rathnayake et al., 2020); however, NDVI is known to be sensitive to soil background and atmospheric effects as well. Time series of Tasseled Cap Angle (TCA), which characterizes percent vegetation cover based on the arctangent of the TCG-TCB ratio, can be a useful transform when incorporating Landsat MSS data into the time series record, as the MSS sensor does not include SWIR bands (Pflugmacher et al., 2012). LandTrendr has also been effectively applied to annual time series of other continuous biophysical variables such as biomass time series (Main-Knorn et al., 2013) and cropland and agricultural land classification probabilities (Dara et al., 2018;Yin et al., 2018).
Assuming a univariate relationship between the change processes of interest and the time series of observed values simplifies interpretation of the results of the segmentation process, and vertices can be applied to other bands/indices such that the timing of changes remains consistent across multiple input features. Extending segmentation results to other bands also enables generation of smoothed values that can be used for multispectral classification. However, different inputs may produce different trajectories (Fig. 5), and leveraging relationships between these results requires additional post-processing, such as secondary classification or "stacking" methods Healey et al., 2018).
While CCDC can be configured for single-band change detection (e. g., Yu et al., 2021), CCDC segmentation is typically performed using multiple spectral bands or indices and univariate tests for each band are combined in order to identify breaks. By considering change in a multispectral context, CCDC is less sensitive to individual inputs and break detection is consistent across inputs and derived indices (Fig. 5). In most cases, CCDC is run on surface reflectance inputs with the Blue band (4.5-5.2 nm) omitted due to strong atmospheric influence (Table 3). CCDC has also been run on derived indices, such as the Normalized Difference Fraction Index (NDFI), which uses an endmember-based unmixing approach to provide enhanced information on fractional cover (Bullock et al., 2020b;Chen et al., 2021), and has been applied to time series from other sensors, such as dual-polarized radar inputs from Sentinel-1 (Shimizu et al., 2019a) and MODIS NDVI (X. Tang et al., 2019). However, when multiple inputs contribute to the change detection process, interpretation of CCDC segmentation results can be less straightforward than those from univariate LandTrendr models. Furthermore, while including harmonic terms at several frequencies and employing a regularization approach enables flexibility and generalization in model fitting, it is important to consider whether temporal patterns in the observation record are well-represented using CCDC's inherently cyclic model formulation. For example, Cohen et al. (2020) found that COLD, which uses a very similar modeling and break detection approach as CCDC, tended to show a greater reliance on NIR-based indices than LandTrendr for forest disturbance detection, potentially due to improved characterization of forest phenology from all available observations using harmonic models.

Additional model parameterization
In addition to spectral and temporal attributes of time series inputs, LandTrendr and CCDC have a number of other tunable parameters (see Tables 4 and 5 for default parameter values for Earth Engine and other key reference implementations). Although both algorithms are generally robust to small parameterization changes, different combinations of parameters can produce different segmentation results and parameter values are often selected using example sites prior to scaling to a full study area. For example, in developing a monitoring system to characterize changes in biomass in the Pacific Northwest, Kennedy et al. (2018a) developed three LandTrendr parameterizations for SWIR1, TCW, and NBR, with best model proportion and p-value parameters increased when using SWIR1 and fewer segments used for TCW. When working with CCDC, the number of observations required to flag a change and the chi-square probability threshold for change detection are perhaps the most critical parameters to be tuned. For example, Brown et al. (2019) dynamically adjusted the consecutive observations parameter based on frequency of time series inputs, whereas Cohen et al. (2020) decreased the chi-square probability from 0.99 to 0.90 to better tolerate commission errors.
Other studies have included more formal sensitivity analyses to optimize segmentation results. For example, Xu et al. (2019) used a parametric tuning approach to systematically select a set of optimal parameters for LandTrendr segmentation and post-processing, and found the greatest variability in values for spikeThreshold, vertex-CountOvershoot, recoveryThreshold, and bestModelProportion across different spectral inputs, specifically NBR, NDVI, NDMI, the Red band, and the SWIR1 band. In the CCDC context, Awty-Carroll et al. (2019b) used a manual interpretation of model results to select a value for the lambda parameter and found that a value of 1 (rather than the default of 20) worked best for their use case.
Unless otherwise noted, the examples presented in this paper were generated using the Earth Engine default parameters, with two notable exceptions. First, the LandTrendr default best model proportion was decreased from 1.25 (Table 4) to 0.75, as the former typically produced very poor results and the latter is more consistent with most published case studies. Second, because we worked with scaled surface reflectance values that were divided by 10000, we also re-scaled our CCDC lambda parameter from 20 (Table 5) to 0.002.

Outputs and data products
Whereas visualizing segmentation results for individual pixels is essential for understanding spectral-temporal relationships and parameterizing algorithms to detect change processes of interest, development of spatial products is typically the goal for most temporal segmentation analyses. The LandTrendr and CCDC temporal segmentation functions available in Google Earth Engine and from other sources (Table 1) operate on a per-pixel basis and typically generate outputs in the form of array images. Unlike conventional rasters that store only one value or observation per band, array images can contain multi-dimensional arrays of values for each pixel, enabling more complex indexing of the spectral-temporal domain. This format lends itself well to temporal segmentation where multi-spectral segment and vertex/break information must be stored for each pixel. For example, CCDC stores the following information per pixel: (a) segment start and end dates and dates of detected breaks, (b) number of observations per segment, (c) change probability per detected change, (d) CCDC coefficients corresponding to intercept, slope and pairs of harmonic sine and cosine terms by number of temporal segments, (e) Root Mean Square Error (RMSE) of each temporal segment for each input spectral band, and (f) magnitude of change for each input band. Similarly, the output from the Land-Trendr algorithm consists of: (a) the year of observations, (b) the source value of observations, (c) the source value of observations fitted to segment lines between vertices (Fit-To-Vertex, or FTV), (d) a Boolean indicator for whether an observation is considered a segmentation vertex, (e) the RMSE of the FTV values, relative to the source values, and (f) complete time series FTV values for additional bands in the collection greater than band 1.
To summarize information and generate more standard map products, array images are subset or "sliced" in the temporal domain to extract segment, vertex, and break attributes that can be sorted and combined in a variety of ways. Although variable-length CCDC arrays are inherently more complex than LandTrendr arrays, which are structured in terms of number of years, both can be used to produce very similar products (Table 6). For example, both algorithms can be used to map the time of first, last and/or largest change over a specified time horizon and provide information on the magnitude of change between two segments (Fig. 6). Both also provide RMSE as a metric of model fit, and segmentation results can be used to generate "synthetic" or fitted spectral values that are essentially smoothed in the temporal domain. These synthetic images can be used in place of observed imagery or composites for change detection and/or land cover classification (e.g., Healey et al., 2018;Kennedy et al., 2018a). CCDC segmentation outputs a set of regression coefficients that can also be used for classification of segments and/or specific time periods (e.g., Brown et al., 2019;Pasquarella et al., 2018;Xian et al., 2021). Additional attributes that can be derived from LandTrendr segmentation results include segment duration and rate of change, pre-change FTV, and a Change Signal-to-Noise Ratio (CSNR), previously referred to as the Disturbance Signal-to-Noise Ratio (DNSR), which is calculated by dividing the change signal by the plot-level noise estimated by the model RMSE .

Tools and visualization
The availability of LandTrendr and CCDC functions in Google Earth Engine has enabled the development of new tools that facilitate running these algorithms and interacting with their outputs. Graphical user interfaces for plotting time series and LandTrendr fitted data and mapping disturbances and visualizing change are available in the UI Applications section of the LT-GEE Guide (emapr.github.io/LT-GEE/). Current tools include Pixel Time Series Plotter, Change Mapper, Fitted Index Delta RGB Mapper, and Time Series Animator interfaces, as well as example scripts for producing maps of vegetation distance, gain, and loss. Similarly, a set of new interactive tools to run and visualize time series of CCDC fitted results, create maps of CCDC coefficients (including transforms such as phase and amplitude) and predicted images, and create land cover and change maps are described in Arévalo et al. (2020a). Both sets of tools also offer their own application programming interface (API) functionality that allows users more flexibility to interact with the algorithm outputs.
To facilitate comparisons like those shown in this study, we have also developed a tool to visualize and compare segmentation results from LandTrendr and CCDC (see Supplemental materials). This tool runs both algorithms on Landsat time series for the same pixel, allowing users to interactively adjust inputs and parameters and view segmentation results. As implementations of CCDC, LandTrendr, and other temporal segmentation algorithms continue to mature, visualization and other pre-and post-processing toolkits will be essential for facilitating adoption by broader remote sensing and applications communities and building a deeper understanding of best practices in terms of use cases and parameterization choices.

Discussion and conclusions
When selecting among a growing suite of temporal segmentation algorithms, users must consider the variety and complexity of land cover change processes of interest, including their temporal duration and expected spectral-temporal trajectories. Though LandTrendr and CCDC are both designed to automate extraction of reflectance characteristics and change event information from time series of satellite imagery, these methods are distinct in their use of the spectral and temporal domains, fitting approaches, and attributes of segmentation results. Both algorithms assume that a sequence of linear models can be used to identify and quantify periods of stability and change in surface properties, and both rely on the availability of sufficient high-quality measurements to  Kennedy et al. (2010) and https://developers.google.com/earth-engine/apidocs/ee-algorithms-temporalsegmentation-landtrendr for parameter descriptions.  Intercept, slope, sine/cosine coefficients V.J. Pasquarella et al. characterize patterns and processes of interest. LandTrendr typically uses univariate time series reduced to an annual time step to generate continuous reflectance trajectories, whereas CCDC uses all clear observations across multiple spectral bands or indices to perform break detection and characterize seasonal variability within segments. A summary of key comparisons is provided in Table 7. Some important considerations for potential users can be broadly described in terms of scalability, sensitivity, reproducibility, and accuracy of algorithms and outputs. Although both the LandTrendr and CCDC algorithms can now be run in Google Earth Engine, users still face notable trade-offs balancing the precision of information generated by the segmentation process and associated storage and runtime. Land-Trendr is less computationally intensive than CCDC and can be applied over large areas with relative ease. CCDC is arguably less scalable because of larger computing and storage requirements, and generating CCDC results at national and continental scales may remain prohibitive for most users. However, CCDC segmentation results include information on seasonal dynamics as well as more precision in estimating the timing of disturbances and other land cover changes that may be critical for some mapping and monitoring applications. For example, Land-Trendr is well suited to providing yearly disturbance estimates, but if sub-yearly information is critical, then CCDC is likely a better choice, assuming a sufficient number of high quality observations are available for the time period(s) of interest.
Users should also consider the sensitivity of the LandTrendr and CCDC algorithms with regard to inputs and parameters. LandTrendr users must make informed choices regarding compositing period and individual spectral indices used for segmentation, and small changes in spectral inputs and key parameters can substantially impact segmentation results. To optimize fitting for particular regions, ecosystems, and change processes, LandTrendr parameter testing is recommended to calibrate inputs and model settings. CCDC users may also adjust a number of model specifications, including input bands used for segmentation and regression tuning parameters, and an exploratory calibration phase is also highly recommended. As a multivariate model that uses all available observations, the CCDC approach is generally not as sensitive to inputs and small parameter changes as LandTrendr; however, CCDC's iterative fitting and change detection approach is more sensitive to initial conditions and changes in observation frequency.
Sensitivity of segmentation results to spectral and temporal inputs and parameterization choices also plays an important role in reproducibility. Because temporal segmentation is highly dependent on a number of user-specified decisions and a given result represents just one of potentially many possible outcomes, thorough documentation of algorithm inputs is essential for ensuring reproducibility of results. Users, particularly those working with Earth Engine implementations, are therefore encouraged to provide complete details on not only spectral and temporal properties of image inputs, but also parameter values and any post-processing decisions. As image archives continue to be updated with new acquisitions, there will also be unprecedented opportunities to revisit previous temporal segmentation studies and establish best practices for updating and improving results.
A final point of consideration involves the accuracies of disturbance maps generated from CCDC and LandTrendr. CCDC and LandTrendr fundamentally differ in their modeling approaches, and therefore users should expect to see differences in resulting land cover and change products. CCDC is well suited for detecting high-magnitude disturbances and land cover changes with minimal tuning across diverse land cover types. However, it has been found to be less suitable for detecting more gradual land cover transitions that are not associated with a distinct disturbance event, such as insect disturbances, partial/selective harvests, and vegetation regrowth (Cohen et al., 2017), though modifications to the core CCDC approach such as COLD (COntinuous monitoring of Land Disturbance) can improve detecting such transitions . LandTrendr targets both discrete and gradual trends and can be fine-tuned to specific processes of interest. Multi-spectral ensembles that combine information from multiple LandTrendr runs can also be used to produce more robust change detection results De Marzo et al., 2021).
Continued maturation of temporal segmentation algorithms and their implementations has also resulted in opportunities to improve segmentation outputs by combining desirable features of each segmentation algorithm. For example, a version of the F-statistic test used by LandTrendr has been integrated as a post-processing step for CCDC results. In this "commission test," breaks that do not significantly improve model fit are eliminated from CCDC change results, resulting in improved land cover and land cover change products . Alternatively, harmonic features can be estimated using fixed window time series (e.g., Adams et al., 2020;Wilson et al., 2018), and associated regression coefficients may be used as inputs to or in conjunction with LandTrendr segmentation results to provide additional information on more complex phenological cycles. Combinations of output maps generated with different algorithms and/or parameterization can also be leveraged to improve the quality of change detection results, and map-to-map comparisons (e.g., Cohen et al., 2017) and other ensembling approaches Healey et al., 2018;Hislop et al., 2019) provide new insights into the relative strengths and weaknesses of segmentation approaches for specific applications. However, ensemble approaches inherently require greater computational resources and data storage, and more rigorous assessment of tradeoffs between gains in accuracy and model complexity and interpretability represents an important direction for future research. With a growing number of temporal segmentation algorithms now implemented natively in Google Earth Engine and being made available by the Earth Engine user community (e.g., Hamunyela et al., 2020), there are new opportunities to test these algorithms individually and in combination for a diverse array of possible applications. By having centralized code repositories and tools paired with versioned image collections available in a cloud-computing environment like Google Earth Engine, the effectiveness of different algorithms for detecting different types of change processes occurring in different types of ecosystems/landscapes can be more formally tested and best practices for selecting and parameterizing different algorithms for specific types of mapping and monitoring can be developed. Furthermore, users will increasingly have a choice between running temporal segmentation and other image pre-and post-processing algorithms themselves or developing applications on top of existing results and other "analysis-ready" feature sets and products. As existing cloud-based datasets and analysis tools continue to mature and evolve and new datasets and tools become available, ongoing comparison of conceptual and methodological differences, such as the CCDC-LandTrendr comparison presented here, will be essential for both informed use and future development of image processing algorithms and their associated products.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.