Introduction

This report documents the methodology and results gained from a multitemporal geospatial analysis of deforestation rates between 2001 and 2021 in the Satipo region of Peru. It includes analysis of deforestation rates within community areas and in the greater region of Satipo (Image 1) and includes an analysis of historical (1980s-early 2000s) deforestation rates for reference. Five primary comparisons are included: (1) a comparison of the deforestation rate within each community to the larger province, Satipo, with protected areas included; (2) a comparison of the deforestation rate within each community to Satipo, with protected areas excluded; (3) a comparison of the deforestation rate within each community to the larger watershed, with protected areas included; (4) a comparison of the deforestation rate within each community to the larger watershed, with protected areas excluded; and (5) an estimate, based on current rates, of hectares that will be deforested in 2022 and 2023 and an estimate of what a 30% reduction in deforestation might look like.

Image 1: Satellite overview of the Kemito Ene River

Methodology

Analyses were performed in Google Earth Engine [1] and R programming language [2].The LandTrendr (Landsat-based detection of Trends in Disturbance and Recovery, [3]) algorithm was used to analyze time-series stacks of Landsat 5-8 data across the study years. The LandTrendr approach allows the user to assess yearly land surface change due to pixel-based spectral changes (Image 2). This type of time-series analysis is desirable for simultaneously detecting short-term changes while smoothing long-term trajectories, and vastly improves on previous change detection methodologies that relied on analysis of two points in time, or on sequences of two-date comparisons. The complexities of multi-temporal land changes - even subtle changes - can be captured and better separated from background noise with the unique LandTrendr method of arbitrary temporal segmentation, which is malleable to the data rather than relying on fixed models of change [3,4,5].

Image 2: Results of a LandTrendr analysis of forest disturbance for the Kemito Region

To ensure the capture of changes in forest cover alone, any pixels of rivers or river disturbances were masked out and excluded from the analysis. A GIS data layer of rivers was downloaded from a global rivers database [6] and used to mask out river pixels based on river location and width (Image 3). As discussed in the Future Recommendations section below, this dataset proved to be imperfect in its coverage of streambeds and seasonal flow paths, leading to the inevitability of some inclusion of forest disturbance attributable to river flow path changes over time. The proportion of these river pixels to the entire study area is small and unlikely to influence the overall conclusions of this analysis.

A challenge in multitemporal geospatial analysis is effectively accounting for uncertainty and propagating that uncertainty through each individual analysis.To faithfully document uncertainty in this analysis, data were categorically split into low, medium, and high magnitude disturbance bins and low, medium, and high DSNR (differential signal-to-noise ratio) bins. Annual disturbance patterns were compared by bin to understand the overall paradigms of disturbance over time. With regard to the uncertainty and accuracy of this method, options are presented in the Future Recommendations section below.To assess change in deforestation over time, linear regression was used to analyze annual deforestation levels across all community areas. Regression lines were fit to data between 2000 and 2021, and between only 2010 to 2021 data. The fit was much improved (0.62 maximum r-squared versus 0.15 maximum r-squared) when the data were constrained to 2010-2021, likely due to management changes at that time. Deforestation rates were extracted from the 2010-2021 regression lines to extrapolate deforestation rates to the future years of 2022 and 2023, and to predict the results of a 30 percent reduction in deforestation.

Results

Broadly, results from LandTrendr show high forest disturbance in Satipo in the 1980s and early 1990s, lower disturbance between 2000 and 2010, and higher disturbance once again between 2010 and 2021 [Fig. 1]. However, paradigms of disturbance differed among these time periods. While disturbances in the Satipo region during the 80s-90s were predominantly characterized by medium and low magnitude events, more recent disturbances were largely characterized by higher magnitude disturbance events.

When comparing individual communities to the larger region of Satipo, the same general forest disturbance patterns are evident, but the magnitude of overall disturbance is less in the larger region than in individual communities. For example, the annual rate of disturbance in the greater Satipo region in the 80s to mid 90s is one to two percent, then drops to around half a percent in the early 2000s. Starting close to 2010, there is a clear upward trend in average forest disturbance to a high near five percent in 2021 [Fig. 1].

Figure 2: Rates of forest disturbance from 1985 to 2021for community groups, broken out by magnitude and differential signal to noise ratio.

At the regional Satipo scale, most disturbance is attributed to low and medium disturbance events, with less contribution from high magnitude events across all years. This is in contrast to results from an analysis of individual communities, which shows a consistent pattern of high magnitude disturbance events in the 80s-90s and again after the year 2000 [Fig. 2].

Changes in rates of deforestation were consistent within groups ('Satipo Community Areas, All','Satipo Minus Protected Areas, All','Satipo, All','Watershed Minus Protected Areas, All','Watershed, All') and similar rates of change were seen across all groups. Regressions fits using all data from 2000 onwards were substantially poorer than regression fits using only data from 2010 onwards. In comparing both the watershed and the broader Satipo region, when protected areas are accounted for we see a very slight decrease in the rate of deforestation.

Figure 3: Best fit lines for linear models based on either post-2000 (dotted lines) or post-2010 (dashed lines) disturbance data

Using data from after 2010 to predict future deforestation resulted in better fit linear models than using data going back as far as the year 2000. When using data as far back as the year 2000, the group with the highest fit for predicting rates of deforestation was 'Satipo Minus Protected Areas, All' (Table 1; r-squared, 0.15). This fit improved substantially by limiting the regression to data after the year 2010 (Table 2; r-squared, 0.62). Improved fit was observed across the board in all groups when compared to limiting the data for predicting deforestation to the period after 2010. The lowest fit category in the post-2010 prediction of deforestation of was the 'Satipo Community Areas, All' (Table 2; r-squared, 0.11). This fit is improved substantially by estimating individual functions for each community, with no community resulting in a poorer fit, and some communities increasing their fit substantially (Table 3; by community minimum r-squared 0.11, maximum r-squared 0.80).

Table 1: Predicted deforestation in hectares using data after 2000

Table 2: Predicted deforestation in hectares using data after 2010

Table 3: Predicted deforestation in hectares using data after 2010 by Community Group

Conclusions

Some community areas demonstrate a far stronger rate of deforestation than others. Based on the wide discrepency with regards to the predicive capacity of a linear regression to predict future deforestation rates, this is strongly indicative of very different land use change patterns amoung the community groups. While some community groups are demonstrating rates of 4-7% deforestation annually, some community groups' deforestation rates are indistinguishable from background rates of forest disturbance and are on par with some of the lowest rates of deforestation observed in the late 90's and early 2000's. Managing and implementing a reduction in deforestation in this regard will require a careful consideration and targeting of specific areas for mitigation.A linear regression may not be the most reliable method for predicting future states of deforestation. Arbitrary inclusion and disclusion of data with regards to what data a model will consider can generate arbitrary fits that may or may not be useful in predicting future rates of deforestation. For example, some community groups which are currently undergoing deforestation have seen some of their most significant increases in deforestation in the very near past. Since both sets of regression models considered in this study have arbitrary cut-offs (necessarily) at some point in the past, this is likely to underestimate the current rates of deforestation. An alternative might be to identify super-regional high rates of deforestation, as well as estimates for a background rates of deforestation, and to recommend target deforestation rates that can reasonably be achieved within this context. Likewise for ongoing monitoring, a more predictive rate of future deforestation might be identified by using the three previous years as a rolling average, and basing a 30% reduction in deforestation on that average. Considering the manner in which previous periods of deforestation have manifested, assuming linear rates of deforestation is likely not the most predictive method for establishing management outcomes.

Figure 4: Best fit lines for community groups

Notes on the Five Comparisons

Although the below comparisons are split into categories that include and exclude protected areas, results do not differ significantly based on protection status.

Five Comparisons:

(1) A comparison of the deforestation rate within each community to the larger province, Satipo, with protected areas included:

Satipo, with protected areas included: The deforestation rate within individual community areas differs. There are four community areas that exhibit particularly high forest disturbance relative to the greater Satipo region, while the majority of community areas exhibit lower disturbance than that of the greater Satipo region.

(2) A comparison of the deforestation rate within each community to the larger province, Satipo, with protected areas excluded:

The same pattern is evident here as with the protected areas included.

(3) A comparison of the deforestation rate within each community to the larger watershed, with protected areas included:

As with the comparison between the community areas and Satipo, a few community areas exhibit greater deforestation rates than those of the larger watershed, while many exhibit lower rates.

(4) A comparison of the deforestation rate within each community to the larger watershed, with protected areas excluded:

Excluding protected areas does not appear to influence the estimated deforestation rates.

(5) An estimate, based on current rates, of hectares that will be deforested in 2022 and 2023 and an estimate of what a 30% reduction in deforestation might look like:

Refer to values in tables 1-3.

Comparison between the current analysis and previous analysis

Over the greater Satipo region, the percent loss annually is generally in agreement. Our analysis differs from the previous analysis in the percentages of annual deforestation in individual communities; we show annual percent loss of up to 11 percent in the late 80s and early 90s, and up to 7.5 percent in 2020 (Fig. 1).

Future Recommendations

The LandTrendr algorithm was developed and parameterized based on Pacific Northwest (Oregon, Washington, and California) forest structure and disturbance. Therefore, it is recommended to further refine the model to capture rain forest disturbance in the tropics by hyper parameterizing the LandTrendr algorithm to features and land management specific to tropical regions. These differences between Pacific Northwest forests and rain forests in tropical regions include forest regeneration rates, seasonality of cloud cover, number of days per year of cloud cover, and management considerations (e.g., harvest methods), among others.A second recommendation is to construct an improved river mask to better exclude disturbance due to seasonal and annual river movement. The mask used in the present analysis was a simple global model of river width with the stated accuracy of 70 percent within the region used for validation, which covered the Columbus River, USA, and did not include any South American rivers for validation. Manually QCing the river extents in the study area is the recommended method for improving the river mask.Field data describing the intensity and frequency of regional disturbance patterns would be the ideal method for establishing an uncertainty regarding the detection rates of disturbance. Alternatively, photo interpretation of high magnitude disturbances is easier and more reliable than photo interpretation of low and moderate magnitude disturbance.

Technical Notes

Due to issues loading certain packages in Deepnote, analyses were run in two separate environments. Owing to the larger nature of some of the spatial data (rasters), a larger than free-tier instance was used for running the 'Kemito_Ene_Processing' script.

Deliverable Details

Tabular data as part of the delivery of this project are in the folder /work/Tabular-Data. Additionally /work/Tabular-Data/ForestLoss_Satipo_Predict.csv has the analysis ready comparisons to relate this work to previous or future work.

References

[1] Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27.

[2] R Core Team (2017). R: A language and environment for statistical computing. R Foundation for Statistical Computing,Vienna, Austria. URL https://www.R-project.org/

[3] Kennedy, R. E., Yang, Z., & Cohen, W. B. (2010). Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr—Temporal segmentation algorithms. Remote Sensing of Environment, 114(12), 2897-2910.

[4] Cohen, W. B., Yang, Z., Healey, S. P., Kennedy, R. E., & Gorelick, N. (2018). A LandTrendr multispectral ensemble for forest disturbance detection. Remote sensing of environment, 205, 131-140.

[5] Kennedy, R. E., Yang, Z., Gorelick, N., Braaten, J., Cavalcante, L., Cohen, W. B., & Healey, S. (2018). Implementation of the LandTrendr algorithm on google earth engine. Remote Sensing, 10(5), 691.

[6] K. Andreadis, G. Schumann, and T. Pavelsky, A simple global river bankfull width and depth database, Water Resour.Res., in review. http://gaia.geosci.unc.edu/rivers/.