Landscape Context for Forest Transition Success in Central Panama

doi:10.21203/rs.3.rs-2594990/v1

Download PDF

Research Article

Landscape Context for Forest Transition Success in Central Panama

https://doi.org/10.21203/rs.3.rs-2594990/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 27 May, 2023

Read the published version in Landscape Ecology →

You are reading this latest preprint version

Context

Secondary forests are frequently re-cleared before they can recover to pre-disturbance conditions. The identification of factors associated with passive regeneration persistence success would help planning cost-efficient forest restoration.

Objectives

In this paper we investigated the role that the landscape context of naturally regenerated forest patches plays for their chances to mature and persist in time in central Panama. Maturation and persistence of secondary forests are concepts often undervalued representing, however, essential requisites for an effective and long-term restoration of the ecosystem processes.

Methods

A unique data set of land-cover maps of central Panama was used to identify the forest patches that naturally recovered and persisted between 1990 and 2020. We developed a Random Forest Classification (RFC) calibration method to identify areas with higher likelihood of forest persistence success.

Results

The RFC model discriminated between areas that naturally recovered and persisted in time and areas that did not persisted with an error rate of 2%. By tuning, we obtained a precision of 0.94 (94%) in the validation test. Based on the model, we developed a prediction map of central Panama areas with higher probability (≥ 90%) of secondary forests persistence success within the next 20 years.

Conclusions

Tracking simple landscape and socio-economic metrics allowed for a deeper understanding of the underlying mechanisms of secondary forest persistence in central Panama. Through the development of RFC calibration method, this study maximized the reliability of the patches identified as suitable to persistence success, representing a basis for management decisions and future investigations for a successful, long-term forest-landscape restoration.

tropical forest-landscape

landscape restoration

landscape analysis

machine learning

Panama Canal Watershed

passive regeneration

human-modified landscapes

persistence

forest transition

Protecting and restoring forests is essential to meeting the Paris climate and the new Global Biodiversity Framework goals, conserving biodiversity and addressing food security and livelihood needs (Chazdon et al. 2017; UN SDG 2017; Lewis et al. 2019; CBD 2022; Hall et al. 2022). Many reforestation activities are being conducted worldwide, at both local and landscape scales. While local-based projects could help determine the best specific solution for the target area, landscape-scale projects help setting a long-term perspective and planning for restoration activities (Mansourian and Vallauri 2014). The Forest Landscape Restoration (FLR) is a holistic approach that considers the relationship between landscape structure (the spatial patterning of different land uses across space) and landscape functioning (ecological processes) and holds promises to achieve both ecological and social goals. It also provides long-term, multi-objective and large-scale means to implement international targets in field interventions(Turner 2005; Hall et al. 2011, 2015; Holl 2017; Stanturf and Mansourian 2020). Passive regeneration of forests plays a major role in large-scale restoration in tropical regions (Chazdon and Guariguata 2016; Sinacore, K. et al. (submitted)). Restoration methods based on passive regeneration are considerably less costly than those based on planting trees, enabling more cost-effective, large-scale forest restoration (Chazdon and Uriarte 2016).

Significant progress in restoring ecosystem services found in mature tropical forests have been observed in two decades of naturally recovering secondary forests (Arroyo-Rodríguez et al. 2009; Hassler et al. 2011; Martin et al. 2013; Rozendaal et al. 2019; Poorter et al. 2021), but newly regenerated forests are frequently re-cleared before they can recover to pre-disturbance conditions (Breugel et al. 2013; Mbow et al. 2017; Meli et al. 2017; Brondizio et al. 2019; Reid et al. 2019). While most studies tend to focus on the socio-economic drivers of deforestation, few provided insights about the conditions needed for secondary forests to grow and persist over time (Smith et al. 2003; Fagan et al. 2013; Sloan 2016; Reid et al. 2017; Schwartz et al. 2017; Borda-Niño et al. 2020). Therefore, there is a strong urgency to identify factors or landscape charachteristics associated with passive restoration persistence success. This would allow to develop predictive models for cost-effective forest-landscape restoration, by favoring areas with higher likelihood of persistence (Schwartz et al. 2020).

Central Panama represents the perfect laboratory to address this question. In this paper, we aim at extending our knowledge of forest transition persistence by studying the forest areas in central Panama that naturally recovered between 1990 and 2010, and persisted until 2020. We used the Random Forest Classification (RFC) and relied on the tuning of important RFC parameters to identify areas with higher likelihood of forest transition success. The tuning process allows Random Forest Classification to be trained accounting for the specific restoration project’s objectives and limitations.

2.1 Study Area

The study region included 23% of the country’s land area and enclosed the mainland area of central Panama (Fig. 1). Over half of the nation’s population lives within the study area (Instituto Nacional de Estadistica y Censo de Panama 2019) which includes the Panama Canal Watershed (PCW), an area of critical economic importance (Adamowicz et al. 2019) where earlier projections of deforestation and carbon storage have been conducted (Dale et al. 2003; Hall et al. 2022a). The region is also important for its provision of and dependence on ecosystem services (Heckadon Moreno et al. 1999; Condit et al. 2001a; Ibanez et al. 2002; Hall et al. 2015). We followed Hall et al. (2022) and included the area outside of the PCW because of its potential role for ecosystem recovery in Panama (Wright and Samaniego 2008). It is a dynamic and heterogeneous area with one of the most diverse ecosystems worldwide, which has experienced rapid economic growth over the past decade, followed by intense land use/land-cover changes. The study area includes the provinces of Panamá and Colón, with a total population of approximately 298,000 residents and hosts one of the most important inland commercial waterways, the Panama Canal, which is a crucial economic driver for the country. It contains two large freshwater lakes – Gatun Lake and Alhajuela Lake – both created by damming for canal operations and freshwater supply for Panama City, Colon City, and numerous towns in between. Until today, the long-term maintenance and sustainability of central Panama biodiversity and water supply depended on the preservation of the tropical forest cover. This worked through a system of protected areas located within two large blocks of land, one in a 16 km corridor along both sides of the former Canal Zone and the other in the sub-basins northeast of Lago Alhajuela, particularly the upper Río Chagres basin.

Yet, the construction of roadways and the expansion of pasture has stimulated the decrease of forest cover capable of regulating water supply and maintaining connectivity between the protected areas (Rompre et al. 2008). The vast majority of the pasture, grasslands, and shrublands are concentrated in between the two protected areas, where only a patchwork of forest fragments is left untouched (Condit et al. 2001b; Stallard et al. 2007).

However, an increase in forest cover has been recorded in central Panama since 1990, caused by the natural expansion of young secondary growth, which offset the deforestation process resulting in no net loss of forest (Walker 2020). This event is called “forest transition”, and it has been increasingly observed in the tropics since the 1990s (Sloan 2022).

2.2 Data Processing and Quantification of Persistence

A unique 30-year time-series data set (1990–2020) of central Panama vegetation cover compiled by Walker (2020) was used to identify areas where natural regeneration occurred spontaneously and persisted between 1990 and 2020 (Fig. 2a, b, c). Before Walker’s work, Panama lacked a consistent dataset by which to assess baselines for land-cover changes. Walker forest cover and forest-cover change maps in Panama were designed to achieve an adequate spatial and temporal resolution. That allows distinguishing different age classes of secondary forests: low vegetation (1–2 yrs), medium vegetation (3–5 yrs), high vegetation (6–20 yrs) and forest (≥ 20 yrs), and therefore capturing small and ephemeral clearings.

Land cover layers were based on epochal composite of least-cloudy, dry-season Landsat images, hand-classified as 5 macro-categories and 23 sub-categories (Table 1). Sub-categories provide additional information regarding habitat, disturbance, and land-use, allowing for a better understanding of land-cover change dynamics. Unlike previous land-cover maps of Panama, which apply dichotomous or categorical definitions of forests, these maps are based on an ordinal scale to define vegetation. Such gradient reflects the natural process of regeneration and acknowledges the fuzzy boundaries between forest and non-forest classes. The regeneration and persistence dynamics were observed working at macro-categories level (NOVEG, LOWVEG, MEDVEG, HIGHVEG and FOREST), while the pre-selection of eligible patches also considered the sub-categories level. We adopted a patch-based approach, where a patch is defined as a minimum 0.5 ha wide, homogeneous area that differs from its surroundings (Forman and Godron 1981; Farina 2000). The patches classified as HIGHVEG in 1990 were excluded from the dataset to limit the fuzziness and highlight the transition from non-forest to forest. The patches classified as NOVEG, as LOWVEG sub-categories low crops rainfed (18) and low crops irrigated (28), and as MEDVEG sub-category high crops (48) were not considered in the analysis as our goal was to assess the forest-landscape passive regeneration potential in a multifunctional landscape, without altering the presence of urban/agricultural land-uses. Finally, we excluded the patches classified as FOREST sub-category evergreen plantations (68) to focus on natural regeneration only. We designated patches with 0 = passive regeneration BUT NO persistence or 1 = passive regeneration AND persistence histories as follows: patches 1 were identified as ≥ 0.5 ha patches classified as LOWVEG or MEDVEG in 1990 land cover map and as FOREST in both subsequent 2010 and 2020 land cover maps, for at least 50% of their total area. Patches 0were identified as ≥ 0.5 ha patches classified as LOWVEG or MEDVEG in 1990, as FOREST in 2010 and again as LOWVEG or MEDVEG in 2020 land-cover map. Our sample was composed by 1,234 patches ranging from 0.5 ha to 35,978 ha, representing a total area of 150,381 ha.

Table 1

Macro and Sub-categories of land-cover maps.
Macro-categories	Summary	Sub-categories: Vegetation		Sub-categories: Wet detail		Sub-categories: Anthropological detail
NOVEG	Water/ Bare/Beach/ Built			-7	Water	1	Built/Bare
NOVEG	Water/ Bare/Beach/ Built			7	Water & Bare	1	Built/Bare
LOWVEG	Low growth (1–2 yrs) / Pasture / Agriculture/ Low wetland	10	Low veg, seasonally dry	27	Low wetland	18	Low crops rainfed
LOWVEG	Low growth (1–2 yrs) / Pasture / Agriculture/ Low wetland	20	Low veg	27	Low wetland	28	Low crops irrigated
MEDVEG	Medium growth (3–5 yrs) / Shrubby crops/ Medium gallery vegetation	30	Thin gallery/edge veg	37	Low to Med gallery	48	High crops
MEDVEG		40	Med veg	47	Med gallery	48	High crops
HIGHVEG	Young high vegetation (6–10 yrs)/ Immature forest (11–20 yrs)	50	High veg	57	High gallery	58	Deciduous plantations
FOREST	Fragmented/disturbed forest/ Forest within 50 m of edge/ Secondary forest, cleared > 20 yrs prior/ Riparian trees (wet or heavily shaded)/ Forested wetland/ Undisturbed upland forest	60	Disturbed forest	67	Mature gallery	68	Evergreen plantations
		70	Forest wetland	77	Mixed water & mangrove
		80	Mature upland forest	87	Gallery in mature forest

2.2.1 Explanatory Variables

While climatic and edaphic factors affect plant germination, growth, and survival, leading variations in the regeneration process across regions, socio-economic factors and landscape context determine where regeneration is permitted to occur and persist, playing a central role in driving successional pathways within human-modified landscapes, such as tropical ones (Jakovac et al. 2021). Nine remote-sensed, easily collectable explanatory variables implying the anthropic use and access of the area were measured for each of the 1,234 patches included in the dataset (Table 2): mean patch elevation (meanElev); mean patch slope (meanSlope); minimum distance from the centroid to the nearest road (RoadsDista); minimum distance from the centroid to the nearest river (Hydro123D); minimum distance from the centroid to the nearest urban area (UrbanDista); minimum distance from the centroid to the nearest rural area (RuralDist); minimum distance from the centroid to the Gatun or Alajuela lake (LLDistDi); minimum distance from the centroid to the nearest protected area (PADistanc); mean distance between the centroid and the patch’s edge (distpole). The selection of explanatory variables was based on a previous study on deforestation scenarios and future carbon dynamics conducted in the PCW (Hall et al. 2022b). The variable distpole, which was not used in the previous study, was included to consider the patch nature of our samples and their variability in terms of extent and shape.

Table 2

Independent variables.
Variables	Dataset name	Description	Source
Slope	meanSlope	Mean Slope in degrees	Digital Elevation Model of the Republic of Panamá, generated by NASA SRTM program, 30m.
Elevation	meanElevation	Mean Elevation in meters	Digital Elevation Model of the Republic of Panamá, generated by NASA SRTM program, 30m.
Distance to Rivers	Hydro123D	Euclidean distance to nearest River (first, second and third order)	Hydrology,Instituto Geográfico Nacional Tommy Guardia, 1:50,000
Distance to Roads	RoadsDista	Euclidean distance to nearest Road	Roads, Open Street Map, 1:50,000
Distance to Protected Areas	PADistanc	Euclidean distance to nearest Protected Area	ANAM (National Environmental Authority, now MiAmbiente), 2006.
Distance to Panama Canal and Alajuela Lake	LLDist	Euclidean distance to nearest Lake (Panama Canal or Alajuela Lake)	Lakes and Lagunes, Instituto Geográfico Nacional Tommy Guardia, 1:50,000
Distance to Urban Areas	UrbanDista	Euclidean distance to nearest Urban Area	Urban Areas, Instituto Geográfico Nacional Tommy Guardia, 1:25,000
Distance to Rural Areas	RuralDista	Euclidean distance to nearest Rural Area	Rural Areas, Instituto Geográfico Nacional Tommy Guardia, 1:25,000
Distance Centroid	distpole	Mean distance between centroid and patch edge	Fragstats 4.2

2.3 Data Analysis

2.3.1 Random Forest Classification

We evaluated the associations between secondary forest persistence success and landscape-context predictors using the Machine-Learning (ML) based method of Random Forest Classification (RFC) (Breiman 2001). The advantage of ML over traditional statistical techniques is the ability to model highly dimensional and non-linear data – such as ecological data (Knudby et al. 2010; Thessen 2016) – with complex interactions. There are several types of tasks that ML techniques can perform. The one task we were interested in was classification, the process of predicting the class of given data point (Kotsiantis 2007). RFC is the most widespread classifier ML technique, especially when working with Land Use/Land Cover Change (LULC) (Vitale et al. 2014; Lowe and Kulkarni 2015; Pelletier et al. 2016; Nguyen et al. 2018; Talukdar et al. 2020). RFC is a supervised algorithm based on decision trees and improved bagging and bootstrap techniques that utilizes several classifiers to work together to identify the class label for each observation (Yang et al. 2010). In addition, RFC provides variable importance measures, which can be used to identify most relevant features and/or their effect (Archer and Kimes 2008). However, like most classifiers RFC faces problems when learning from an extremely unbalanced training data set (Zakariah 2014). As it is constructed to minimize the overall error rate, it will tend to focus more on the prediction accuracy of the majority class, which often results in poor accuracy for the minority class. However, there are several techniques that can be used with RFC to solve class imbalance problems (Chen 2004; Kuhn and Johnson 2013; More and Rana 2017). Here we decided to solve the class unbalance problems to the training set only. We then validated the model on the unbalanced test set, to assess the performance of the model with real frequencies. The study used the randomForest package running in Rstudio (CRAN) (Liaw and Wiener 2002).

2.3.1.1 Training Dataset Balancing

The dataset was composed of 0 = 182 and 1 = 1,052 units. We divided the dataset into training and test sets following the rule of 70%-30%, as it is demonstrated to represent the ratio with the best performance in ML models (Nguyen et al. 2021). The balance problem was solved on the training set through up-sampling technique (upSample function in the caret package running in Rstudio (CRAN)). Up-sampling is an oversampling technique that randomly selects and duplicates observations in the minority class to balance the dataset. We obtained a balanced training set with a total of 1,474 samples. The optimal number of decision trees (n = 1000) and number of variables randomly sampled as candidates at each split (n = 3) were evaluated based on the lowest Out Of Bag (OOB) error. The RFC is trained through bagging technique, where each new decision tree is fit from a subsample of the training units. The OOB error is the average classification error of the remaining training units not included in the subsample. That allows to fit and also validate the RFC while being trained (Breiman 1996).

2.3.1.2 Tuning

The tuning consists in the identification of a set of optimal RFC hyperparameters’ values to maximize the model’s performance and produce better outputs. Hyperparameters are parameters whose values control the learning process of the model and can be tuned to increase some specific metrics like accuracy or precision. In our case, the RFC model was validated on an unbalanced test set to assess the real frequencies of 0 and 1 found in the PCW, within the past 30 years. Since RFC is designed to minimize the overall error rate, with an unbalanced dataset it will tend to focus more on the prediction accuracy of the majority class, which often results in poor accuracy for the minority class. For this reason, accuracy was not a suitable measure for our model’s performance. We instead focused on the tuning of precision for 1, and recall for 0 (Table 3a, b). Here precision is defined as the ability of a classifier not to label “as TRUE” an instance that is actually “FALSE”. For each class, it is defined as the ratio of real trues (RT) to the sum of real trues and false trues (FT), Eq. (1):

$$Precision=\frac{RT}{RT+FT}$$

Table 3. Confusion matrix parameters for class 1 (a.) and 0 (b.). RF=Real False, NRF= Non-Real False, RT= Real True, NRT= Non-Real True.

Confusion matrix 1	Real
Predicted	0	1	Parameters
0	RF	FF	Precision: RT/RT+FT
1	FT	RT	Recall: RT/RT+FF

Confusion matrix 0	Real
Predicted	0	1	Parameters
0	RT	FT	Precision: RT/RT+FT
1	FF	RF	Recall: RT/RT+FF

In other terms, the precision measures how good the model is at predicting a specific category. The recall (also known as sensitivity) is the ability of a classifier to find all the RT instances for the specified class. It is defined as the ratio of RT to the sum of RT and false false (FF), Eq. (2):

$$Recall=\frac{RT}{RT+FF}$$

Recall is a measure of how many times the model was able to detect a specific category. To maximize the precision of 1, we needed to improve the Recall of 0. In other words, we needed to reduce as much as possible the misclassification errors (FT) of 0. A straightforward approach to improve the performance of a classifier that predicts probabilities with unbalanced dataset is to tune the threshold used to map probabilities to class labels. ML algorithms predicts the probability or scoring of class membership by using a threshold. All values equal to or greater than the selected threshold are mapped to one class and all the other values are mapped to another class. We, therefore, tested several thresholds until identifying 0.90 as the optimal one to improve the precision of 1 and recall of 0, as described above. Scores above 0.90 were classified as 1, those below as 0.

However, tuning one of these quantities entailed a trade-off: as we increase the precision of a class, we decrease the recall for the same class, and viceversa. In our case, this trade-off leads to low sensitivity for class 1 (several 1 RT were lost), and consequently low precision for class 0 (a high percentage of the total patches that were predicted as 0, were actually 1).

3.1 Secondary Forest Persistence Analysis

From a total of ~ 62,203,260 ha of LOWVEG and MEDVEG occurring in 1990 in the study area, ~ 150,380 ha (0,24%) were classified as FOREST (passive regeneration) in 2010. Within these, about ~ 36,400 ha (24%) were cleared again by 2020, while ~ 113.900 ha (76%) persisted.

3.2 Random Forest Classification Performance Analysis

As shown in Table 4, in the training phase the RFC model achieved to discriminate between the two classes (0–1) with an OOB error rate of about 2% and a class specific error of 0.001 (0.1%) for the 0, and 0.046 (4.6%) for the 1. By tuning, Precision of 1 and Recall of 0 reached, respectively, 0.94 (94%) and 0.83 (83%) in the validation test (Table 5).

Table 4. Random Forest Classification Training performance.

*Number of decision trees (ntree):*	1000
*Number of variables tried at each split (mtry):*	3
OOB estimate of error rate:	2,37%
Confusion matrix:	Real		Class error
Predicted	0	1	Class error
0	736	1	0.001356852
1	34	703	0.046132972

Table 5. Random Forest Classification Test performance.

Accuracy:	0.5122
95% CI:	(0.4599 , 0.5643)
Mcnemar’s Test P-Value:	<2e-16
Confusion matrix:	Real		Precision	Recall
Predicted	0	1	Precision	Recall
0	45	171	0.46	0.83
1	9	144	0.94	0.20

3.3 Role and Importance of independent variables

The order of importance of the variables was measured through Mean Decrease in Gini (MDG). MDG is a measure of how important a variable is for estimating the class across all of the trees that make up the forest (Grömping 2009). A higher MDG indicates higher variable importance. The most important variables to the model will have the largest MDG. Conversely, the least important variable will have the smallest MDG values. Variables are sorted and displayed in the Variable Importance plot (Fig. 3). The most important variable was meanElev, followed by RuralDist, LLDistDi, UrbanDist, meanSlope, PADistanc, RoadsDista, distpole and Hydro123D. Figure 4 (a, b, c, d, e, f, g, h, i) shows the variables-specific effect on the persistence of secondary forests. The variables meanElev, meanSlope, RoadsDista, UrbanDist and RuralDist appeared to be directly proportional to the probability of 1, while distpole, PADistanc and LLDistDi were inversely proportional to the probability of 1. Of particular interest was Hydro123D, which appeared to be inversely proportional to the probability of 1 within the first ~ 1000 m, and changed to directly proportional as it increased.

3.4 Prediction Map of Secondary Forests Persistence

Based on the RFC model, we developed a prediction map of central Panama areas that have higher probability (≥ 90%) of secondary forests persistence success within the next 20 years (Fig. 5). From a total of 286,593 ha of LOWVEG and MEDVEG present in our study area, only 16% (~ 48,179 ha) have a probability ≥ 0.90% to recover and persist until, at least, 2040.

Natural tropical regeneration has been increasingly recognized as an emergent phenomenon in Latin America. As shown in Walker (2020), central Panama is a hotspot for spontaneous forest transitions. However, our analysis of native secondary forest persistence over the past 30 years revealed that forest transition did not translate into a persistent trend. In other words, many secondary forests were not allowed to mature. This has relevant consequences in terms of climate change mitigation, biodiversity, and ecosystem service recovery. Most of the simulations on passive regeneration for carbon sequestration, biodiversity recovery and/or other desired benefits projections do not account for the effective persistence success of the forests, severely overestimating the potential rate of forest-landscape ecosystem services recovery and climate change mitigation supported by passive regeneration. Van Breugel et al. (2013) demonstrated that this ephemerality could potentially translate into secondary forests’ limited role as biodiversity reservoirs. The proportion of species that can sustain reproductive populations (effective diversity) in a dynamic patchwork of young secondary forests depends on how old the secondary forests become and how many tree species will be able to reach reproductive size within that time frame. Re-clearance of native secondary forests greatly limited carbon sequestration in the Atlantic Forests of Brazil, where second-growth forests could have sequestered over three times more carbon than the actual estimated carbon sequestration (Piffer et al. 2022). Without permanence, secondary forests remain carbon sinks, but the contribution of reforestation to climate mitigation will be severely reduced. Identifying conditions that encourage secondary forest persistence is crucial to have a more accurate idea of the real potential for ecosystem restoration within dynamic landscapes as those found in the tropics. Also, it allows for an informed evaluation of the specific supporting actions needed to reach the UN goals and to ensure a sustained ecosystem services provision.

Tracking simple landscape and socio-economic metrics allowed for a deeper understanding of the underlying mechanisms of native secondary forest persistence (deforestation reversal) in our central Panama study area, between 1990 and 2020. We determined that deforestation reversal was not stochastic. Rather, forests’ persistence could be predicted based on landscape and socio-economic context. Variable importance analysis showed that all the variables used in this study were involved in secondary forests’ persistence success.

Native secondary forests persisted longer in patches that were closer to Gatun Lake, to Alajuela lake and to protected areas, but further from rural communities, roads, urban areas and in patches with higher elevation and steeper slopes, which is in accordance to previous findings within the tropics (Aide et al. 2013, 2019; Molin et al. 2017; Calaboni et al. 2018). Particularly interesting is the role that rural areas have for the persistence success of native secondary forests. Panama has a unique legislation, called Rights of Possession (ROP), which allows people to possess government owned land as long as they manage it and/or make visible improvements upon it to claim ownership. This could explain why in rural areas secondary forest persistence has lower chances to success and underlies the necessity of forest landscape restoration activities to be supported by informed and oriented policies (Slobodian et al. 2020). Contrary to what found in Costa Rica (Reid et al. 2019), where secondary forest patches persisted for longer periods of time when they were larger and close to rivers, in the PCW the secondary forests patches with smaller extents and further from rivers persisted longer. This has potential implications for the passive restoration approach. Small patches can be used as stepping-stones to increase the landscape’s connectivity and facilitate interpatch movements in fragmented landscapes. A study conducted on the tropical rainforest of Los Tuxtlas, Mexico (Arroyo-Rodríguez et al. 2009) showed that, despite the small size of the forest patches, these can potentially contain diverse communities of native plants, including endangered and economically important species. However, the relationship between species’ density and area may vary significantly across landscapes with different deforestation levels. Higher deforestation means positive and significant relationship between species density and patch area. When the dispersal is limited, the neighboring forests will be less floristically diverse and tropical forest recovery will be limited (Holl et al. 2000; Martensen et al. 2012; van Breugel et al. 2019). This highlights the inability of highly dynamic and fragmented landscapes to rely solely on spontaneous forest transition. To foster forest recovery, regions with highly fragmented landscapes should consider facilitating incentive systems for forest-based ecosystem services (Fenichel et al. 2019) and include other specific forest landscape restoration strategies, such as supported reforestation.

Patches further away from rivers persisted longer. This could be explained by the adoption of traditional cattle pasture (CP) in our study area (Stefanski et al. 2015). In this system, cattle have full access along all stream channels. The Panama Canal Authority is promoting silvo-pastoral (SP) system in the PCW, where riparian forest completely surrounds the streams and is protected with fencing to restrict cattle access to a single point. Nevertheless, even though biodiversity would benefit from it, this system could lead to degraded water quality in dry season, caused by increased congregation of cattle to drink and avoid heat through shade (Chavarria et al. 2021). This could be mitigated by the provision of additional tree shade, further away from the streams.

Spontaneous forest transition could, therefore, be crucial for effective ecosystem restoration. However, we found that it may need support via other targeted interventions. A more nuanced understanding of the drivers of forest regeneration persistence in the tropics is critical to ensure the success of forest landscape reforestation efforts, to achieve biodiversity conservation and carbon sequestration targets. Through the development of RFC calibration method, this study maximized the reliability (95% of accuracy) of the patches identified as more suitable to persistence success, representing a basis for management decisions and future investigations for a successful, long-term forest-landscape restoration.

Acknowledgements

This work is a contribution of the Agua Salud Project, a collaboration between the Smithsonian Tropical Research Institute (STRI), the Panama Canal Authority (ACP) and the Ministry of the Environment of Panama (MiAmbiente). Agua Salud is part of the Smithsonian Institution Forest Global Earth Observatory (ForestGEO). We thank Prof. Giovanna Jona Lasinio from Sapienza University, Department of Statistical Sciences who provided insight and expertise that greatly assisted the research. We thank Adriana Tapia, Mario Bailon, and Miguel Nunez for help in ground truthing. We are also immensely grateful to Martina Perez, Danilo Lombardi (Sapienza University, Department of Environmental Biology) and Massimo Pulejo (New York University, Department of Politics) for comments that greatly improved the manuscript. Any errors are our own and should not tarnish the reputations of these esteemed persons.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Competing Interests

The authors have no relevant financial or non-financial interests to disclose.

Author Contributions Statements

G.B. conceived of the presented idea; developed the theory; conceived, planned and conducted the data collection; wrote the manuscript. K.W. contributed with land cover maps. G.B. and G.D.F. verified the analytical methods and performed the statistical analysis. M.V. contributed to writing correction. J.S.H. contributed to improving the manuscript, supervised the findings of this work and encouraged G.B. to investigate on passive regeneration.

Open Research statement

Walker, K. L. 2021. Panama Vegetation Time Series Maps. Figshare data repository. DOI: 10.6084/m9.figshare.14120603. Available at: https://figshare.com/articles/dataset/Panama_Vegetation_Time_Series_Maps/14120603 with supplementary 2020 dataset at: https://figshare.com/articles/dataset/PVCTS2020v2_part_tif/22046813. Shapefiles of Panama topography, roads, populated places, rivers and lakes can be found at STRI GIS Data Portal: https://stridata-si.opendata.arcgis.com/

Conflict of Interest Statement

This study did not collect human subject data nor did it include the study of animals. No authors have any conflict of interest associated with funders or any aspect of data collection.

Adamowicz W, Calderon-Etter L, Entem A, et al (2019) Assessing ecological infrastructure investments. Proceedings of the National Academy of Sciences 116:5254–5261. https://doi.org/10.1073/pnas.1802883116
Aide TM, Clark ML, Grau HR, et al (2013) Deforestation and Reforestation of Latin America and the Caribbean (2001–2010). Biotropica 45:262–271. https://doi.org/10.1111/j.1744-7429.2012.00908.x
Aide TM, Grau HR, Graesser J, et al (2019) Woody vegetation dynamics in the tropical and subtropical Andes from 2001 to 2014: Satellite image interpretation and expert validation. Global Change Biology 25:2112–2126. https://doi.org/10.1111/gcb.14618
Archer KJ, Kimes RV (2008) Empirical characterization of random forest variable importance measures. Computational Statistics & Data Analysis 52:2249–2260
Arroyo-Rodríguez V, Pineda E, Escobar F, Benítez-Malvido J (2009) Value of small patches in the conservation of plant-species diversity in highly fragmented rainforest. Conserv Biol 23:729–739. https://doi.org/10.1111/j.1523-1739.2008.01120.x
Borda-Niño M, Meli P, Brancalion PH (2020) Drivers of tropical forest cover increase: A systematic review. Land Degradation & Development 31:1366–1379
Breiman L (1996) OUT-OF-BAG ESTIMATION
Breiman L (2001) Random Forests. Machine Learning 45:5–32. https://doi.org/10.1023/A:1010933404324
Breugel M van, Hall JS, Craven D, et al (2013) Succession of Ephemeral Secondary Forests and Their Limited Role for the Conservation of Floristic Diversity in a Human-Modified Tropical Landscape. PLOS ONE 8:e82433. https://doi.org/10.1371/journal.pone.0082433
Brondizio ES, Settele J, Díaz S, Ngo HT (2019) Global assessment report on biodiversity and ecosystem services of the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services
Calaboni A, Tambosi L, Igari A, et al (2018) The forest transition in São Paulo, Brazil: historical patterns and potential drivers. Ecology and Society 23:. https://doi.org/10.5751/ES-10270-230407
CBD A New Global Framework for Managing Nature Through 2030: First Detailed Draft Agreement Debuts (2022).
Chavarria KA, Saltonstall K, Vinda J, et al (2021) Land use influences stream bacterial communities in lowland tropical watersheds. Sci Rep 11:21752. https://doi.org/10.1038/s41598-021-01193-7
Chazdon RL, Brancalion PH, Lamb D, et al (2017) A policy-driven knowledge agenda for global forest and landscape restoration. Conservation Letters 10:125–132
Chazdon RL, Guariguata MR (2016) Natural regeneration as a tool for large-scale forest restoration in the tropics: prospects and challenges. Biotropica 48:716–730
Chazdon RL, Uriarte M (2016) Natural regeneration in the context of large-scale forest and landscape restoration in the tropics. Biotropica 48:709–715
Chen C. Using Random Forest to Learn Imbalanced Data (2004).
Condit R, Robinson WD, Ibáñez R, et al (2001a) The Status of the Panama Canal Watershed and Its Biodiversity at the Beginning of the 21^st Century: Long-term ecological studies reveal a diverse flora and fauna near the Panama Canal, harbored within a corridor of forest stretching from the Caribbean to the Pacific, but deforestation, land degradation, erosion, and overhunting remain threats. BioScience 51:389–398. https://doi.org/10.1641/0006-3568(2001)051[0389:TSOTPC]2.0.CO;2
Condit R, Robinson WD, Ibáñez R, et al (2001b) The Status of the Panama Canal Watershed and Its Biodiversity at the Beginning of the 21^st Century: Long-term ecological studies reveal a diverse flora and fauna near the Panama Canal, harbored within a corridor of forest stretching from the Caribbean to the Pacific, but deforestation, land degradation, erosion, and overhunting remain threats. BioScience 51:389–398. https://doi.org/10.1641/0006-3568(2001)051[0389:TSOTPC]2.0.CO;2
Dale V, Brown S, Calderón M, et al (2003) Estimating baseline carbon emissions for the Eastern Panama Canal watershed. Mitigation and Adaptation Strategies for Global Change 8:323–348
Fagan ME, DeFries RS, Sesnie SE, et al (2013) Land cover dynamics following a deforestation ban in northern Costa Rica. Environmental Research Letters 8:034017
Farina A (2000) Landscape Ecology in Action. Springer Netherlands, Dordrecht
Fenichel EP, Adamowicz W, Ashton MS, Hall JS (2019) Incentive Systems for Forest-Based Ecosystem Services with Missing Financial Service Markets. Journal of the Association of Environmental and Resource Economists 6:319–347. https://doi.org/10.1086/701698
Forman RTT, Godron M (1981) Patches and Structural Components for A Landscape Ecology. BioScience 31:733–740. https://doi.org/10.2307/1308780
Grömping U (2009) Variable Importance Assessment in Regression: Linear Regression versus Random Forest. The American Statistician 63:308–319. https://doi.org/10.1198/tast.2009.08199
Hall J, Moss D, Stallard R, et al (2015) Managing Watersheds for Ecosystem Services in the Steepland Neotropics
Hall JS, Ashton MS, Garen EJ, Jose S (2011) The ecology and ecosystem services of native trees: Implications for reforestation and land restoration in Mesoamerica. Forest Ecology and Management 261:1553–1557
Hall JS, Plisinski JS, Mladinich SK, et al (2022a) Deforestation scenarios show the importance of secondary forest for meeting Panama’s carbon goals. Landscape Ecology 37:673–694
Hall JS, Plisinski JS, Mladinich SK, et al (2022b) Deforestation scenarios show the importance of secondary forest for meeting Panama’s carbon goals. Landscape Ecology 37:673–694. https://doi.org/10.1007/s10980-021-01379-4
Hassler SK, Zimmermann B, van Breugel M, et al (2011) Recovery of saturated hydraulic conductivity under secondary succession on former pasture in the humid tropics. Forest Ecology and Management 261:1634–1642
Heckadon Moreno S, Ibañez D. R, Condit R (eds) (1999) La Cuenca del Canal: deforestación, contaminación y urbanización: Proyecto de Monitoreo de la Cuenca del Canal de Panama (PMCC): sumario ejecutivo del informe final. Instituto Smithsonian de Investigaciones Tropicales, Balboa, Ancón, República de Panamá
Holl KD (2017) Research directions in tropical forest restoration. Annals of the Missouri Botanical Garden 102:237–250
Holl KD, Loik ME, Lin EHV, Samuels IA (2000) Tropical Montane Forest Restoration in Costa Rica: Overcoming Barriers to Dispersal and Establishment. Restoration Ecology 8:339–349. https://doi.org/10.1046/j.1526-100x.2000.80049.x
Ibanez R, Condit R, Angehr G, et al (2002) An Ecosystem Report on the Panama Canal: Monitoring the Status of the Forest Communities and the Watershed. Environmental monitoring and assessment 80:65–95. https://doi.org/10.1023/A:1020378926399
Instituto Nacional de Estadística y Censo (2019). https://www.inec.gob.pa/publicaciones/Default2.aspx?ID_CATEGORIA=15&ID_SUBCATEGORIA=60. Accessed 10 Feb 2023
Jakovac CC, Junqueira AB, Crouzeilles R, et al (2021) The role of land-use history in driving successional pathways and its implications for the restoration of tropical forests. Biol Rev Camb Philos Soc 96:1114–1134. https://doi.org/10.1111/brv.12694
Knudby A, Brenning A, LeDrew E (2010) New approaches to modelling fish–habitat relationships. Ecological Modelling 221:503–511
Kotsiantis SB Supervised Machine Learning: A Review of Classification Techniques (2007)
Kuhn M, Johnson K (2013) Applied Predictive Modeling. Springer New York, New York, NY
Lewis SL, Wheeler CE, Mitchard ET, Koch A (2019) Restoring natural forests is the best way to remove atmospheric carbon. Nature 568:25–28
Liaw A, Wiener M (2002) Classiﬁcation and Regression by randomForest. 2:
Lowe B, Kulkarni A (2015) Multispectral Image Analysis Using Random Forest. International Journal on Soft Computing 6:1–14. https://doi.org/10.5121/ijsc.2015.6101
Mansourian S, Vallauri D (2014) Restoring forest landscapes: important lessons learnt. Environmental Management 53:241–251
Martensen AC, Ribeiro MC, Banks-Leite C, et al (2012) Associations of forest cover, fragment area, and connectivity with neotropical understory bird species richness and abundance. Conserv Biol 26:1100–1111. https://doi.org/10.1111/j.1523-1739.2012.01940.x
Martin PA, Newton AC, Bullock JM (2013) Carbon pools recover more quickly than plant biodiversity in tropical secondary forests. Proceedings of the Royal Society B: Biological Sciences 280:20132236. https://doi.org/10.1098/rspb.2013.2236
Mbow H-OP, Reisinger A, Canadell J, O’Brien P (2017) Special Report on climate change, desertification, land degradation, sustainable land management, food security, and greenhouse gas fluxes in terrestrial ecosystems (SR2). Ginevra, IPCC 650:
Meli P, Holl KD, Benayas JMR, et al (2017) A global review of past land use, climate, and active vs. passive restoration effects on forest recovery. PLOS ONE 12:e0171368. https://doi.org/10.1371/journal.pone.0171368
Molin P, Gergel S, Soares-Filho B, Ferraz S (2017) Spatial determinants of Atlantic Forest loss and recovery in Brazil. Landscape ecology
More AS, Rana DP (2017) Review of random forest classification techniques to resolve data imbalance. 2017 1^st International Conference on Intelligent Systems and Information Management (ICISIM) 72–78. https://doi.org/10.1109/ICISIM.2017.8122151
Nguyen HTT, Doan TM, Radeloff V (2018) APPLYING RANDOM FOREST CLASSIFICATION TO MAP LAND USE/LAND COVER USING LANDSAT 8 OLI. In: The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Copernicus GmbH, pp 363–367
Nguyen Q, Ly H-B, Lanh H, et al (2021) Influence of Data Splitting on Performance of Machine Learning Models in Prediction of Shear Strength of Soil. Mathematical Problems in Engineering 2021:. https://doi.org/10.1155/2021/4832864
Pelletier C, Valero S, Inglada J, et al (2016) Assessing the robustness of Random Forests to map land cover with high resolution satellite image time series over large areas. Remote Sensing of Environment 187:156–168. https://doi.org/10.1016/j.rse.2016.10.010
Piffer PR, Calaboni A, Rosa MR, et al (2022) Ephemeral forest regeneration limits carbon sequestration potential in the Brazilian Atlantic Forest. Global Change Biology 28:630–643. https://doi.org/10.1111/gcb.15944
Poorter L, Rozendaal DMA, Bongers F, et al (2021) Functional recovery of secondary tropical forests. Proc Natl Acad Sci U S A 118:e2003405118. https://doi.org/10.1073/pnas.2003405118
Reid JL, Fagan ME, Lucas J, et al (2019) The ephemerality of secondary forests in southern Costa Rica. Conservation Letters 12:e12607
Reid JL, Wilson SJ, Bloomfield GS, et al (2017) How Long Do Restored Ecosystems Persist? 1. Annals of the Missouri Botanical Garden 102:258–265
Rompre G, Robinson DW, Desrochers A (2008) Causes of habitat loss in a Neotropical landscape: The Panamá Canal corridor
Rozendaal DMA, Bongers F, Aide TM, et al (2019) Biodiversity recovery of Neotropical secondary forests. Sci Adv 5:eaau3114. https://doi.org/10.1126/sciadv.aau3114
Schwartz NB, Aide TM, Graesser J, et al (2020) Reversals of reforestation across Latin America limit climate mitigation potential of tropical forests. Frontiers in forests and global change 3:85
Schwartz NB, Uriarte M, DeFries R, et al (2017) Land-use dynamics influence estimates of carbon sequestration potential in tropical second-growth forest. Environmental Research Letters 12:074023
Sinacore, K. et al. (submitted). "Mixed forests success for carbon payments and subsidies to support forest restoration in the neotropics". Nature Communications
Sloan S (2016) Tropical forest gain and interactions amongst agents of forest change. Forests 7:55
Sloan S (2022) Reforestation reversals and forest transitions. Land Use Policy 112:105800. https://doi.org/10.1016/j.landusepol.2021.105800
Slobodian L, Vidal A, Saint-Laurent C Policies that support forest landscape restoration
Smith J, Ferreira S, van de Kop P, et al (2003) The persistence of secondary forests on colonist farms in the BrazilianAmazon. Agroforestry Systems 58:125
Stallard R, Elsenbeer H, Ogden F, Hall J (2007) The Agua Salud Project, Central Panama. AGU Fall Meeting Abstracts
Stanturf JA, Mansourian S (2020) Forest landscape restoration: state of play. Royal Society open science 7:201218
Stefanski SF, Shi X, Hall JS, et al (2015) Teak–cattle production tradeoffs for Panama Canal Watershed small scale producers. Forest Policy and Economics 56:48–56
Talukdar S, Singha P, Mahato S, et al (2020) Land-Use Land-Cover Classification by Machine Learning Classifiers for Satellite Observations—A Review. Remote Sensing 12:1135. https://doi.org/10.3390/rs12071135
Thessen A (2016) Adoption of Machine Learning Techniques in Ecology and Earth Science. One Ecosystem 1:e8621. https://doi.org/10.3897/oneeco.1.e8621
Turner MG (2005) Landscape ecology: what is the state of the science? Annu Rev Ecol Evol Syst 36:319–344
Van Breugel, M., Hall J.S., Craven, D., Bailon, M., Hernandez, A., Abbene, M., Van Breugel, P., 2013. Succession of ephemeral secondary forests and their limited role for the conservation of floristic diversity in a human-modified tropical landscape. PLoSONE.8: e82433
Van Breugel M, Craven D, Lai HR, et al (2019) Soil nutrients and dispersal limitation shape compositional variation in secondary tropical forests across multiple scales. Journal of Ecology 107:566–581. https://doi.org/10.1111/1365-2745.13126
Vitale M, Proietti C, Cionni I, et al (2014) Random forests analysis: A useful tool for defining the relative importance of environmental conditions on crown defoliation. Water, Air, and Soil Pollution 225:. https://doi.org/10.1007/s11270-014-1992-z
Walker K (2020) Capturing ephemeral forest dynamics with hybrid time-series and composite mapping in the Republic of Panama. International Journal of Applied Earth Observation and Geoinformation 87:102029
Wright S, Samaniego M (2008) Historical, Demographic, and Economic Correlates of Land-Use Change in the Republic of Panama. Ecology and Society 13:. https://doi.org/10.5751/ES-02459-130217
Yang P, Hwa Yang Y, B. Zhou B, Y. Zomaya A (2010) A Review of Ensemble Methods in Bioinformatics. CBIO 5:296–308. https://doi.org/10.2174/157489310794072508
Zakariah M (2014) Classification of large datasets using Random Forest Algorithm in various applications: Survey. 4.

No competing interests reported.

Download PDF

Journal Publication

published 27 May, 2023

Read the published version in Landscape Ecology →

Editorial decision: Major revision
06 Apr, 2023
Reviews received at journal
13 Mar, 2023
Reviewers agreed at journal
05 Mar, 2023
Reviewers invited by journal
27 Feb, 2023
Editor assigned by journal
18 Feb, 2023
Submission checks completed at journal
18 Feb, 2023
First submitted to journal
16 Feb, 2023

You are reading this latest preprint version

Landscape Context for Forest Transition Success in Central Panama

Status:

Journal Publication

Version 1

Abstract

Figures

1. Introduction

2. Materials And Methods

2.1 Study Area

2.2 Data Processing and Quantification of Persistence

2.2.1 Explanatory Variables

2.3 Data Analysis

2.3.1 Random Forest Classification

2.3.1.1 Training Dataset Balancing

2.3.1.2 Tuning

3. Results

3.1 Secondary Forest Persistence Analysis

3.2 Random Forest Classification Performance Analysis

3.3 Role and Importance of independent variables

3.4 Prediction Map of Secondary Forests Persistence

4. Discussion

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1