The influence of topography in remote sensing and in the monitoring of forests and environmental services throughout the time

Get Complete Project Material File(s) Now! »

Predictive models to evaluate forest risk loss

Worldwide and in Costa Rica, most of the evaluations targeting the efficiency and additionality of forest conservation policies on forest cover, such as the establishment of PWAs or the PES program have been made using matching techniques (Pfaff et al., 2009, 2014; Robalino et al., 2011; Robalino and Pfaff, 2013; Sierra and Russman, 2006; Wolf et al., 2021). Although these techniques offer reliable results, still lack the incorporation of other factors that can bias the interpretation of the results. They do not incorporate for example if the non-PES farms were eligible or not to participate, or social aspects such as access to information about the program or the effect of tourism as a driver of forest conservation (Daniels et al., 2010). Matching techniques, compared similar groups of farms or PWAs with areas of similar characteristics, but that sometimes were far in location (i.e., across different provinces), lacking important information that can affect socio-economic factors that influence deforestation rates at the local level (Pfaff et al., 2008; Robalino et al., 2011; Wolf et al., 2021). In addition, most of these studies have employed historical deforestation rates, or focused on a short period of time, and projected them forward in a linear fashion, which may introduce biases on predictions, since deforestation rates may respond to complex non-linear variables in time and space (Daniels et al., 2010).
A thorough analysis of deforestation trends and future projections in Costa Rica using biophysical and socioeconomic variables predicted a continuous increase in forest cover, even in the most adverse scenarios, this, at least, demands a better approach of environmental policies for the protection of forests (Stan and Sanchez-Azofeifa, 2019). Analysis like this highlights the importance of identifying areas at higher risk of future deforestation according to patterns observed in historical deforestation trends. In this sense, targeting areas to avoid deforestation will be much improved (Aguilar-Amuchastegui et al., 2014).
The prediction of deforestation risk has been carried out through the use of simple ordinary least square regressions (OLS) in the humid tropics and the Amazon basin, to more complex models such as binary logistic regression in India (Bera et al., 2020), or maximum entropy (MaxEnt) in the Peru Amazon forests (Aguilar-Amuchastegui et al., 2014; Redo et al., 2012). Generalized linear models and generalized linear mixed models (GLMMs), Bayesian networks, and artificial neural networks were compared for the prediction of deforestation in forests in Mexico and Madagascar (Mayfield et al., 2017), and Spatio-temporal Bayesian Network approaches evaluated deforestation risk analysis in Brazil (Silva et al., 2020).
Lately used Machine Learning techniques such as Random Forest (RF) have shown very good accuracies, and in some cases, some advantages compared to previous methods (Breiman, 2001). RF has been implemented for the assessment of deforestation trends and its main drivers in Bolivia (Redo et al., 2012), for predicting future deforestation risk in Borneo (Cushman et al., 2017), and for the spatial prediction of deforestation probability in India (Saha et al., 2020). RF has also shown to be accurate in other ecological disciplines involving forests, such as prediction of tree species presence in the United States (Evans and Cushman, 2009), prediction of forest loss due to wind damage in Southern France (Hart et al., 2019) or susceptibility to landslides in protected and non-protected forests in Iran (Shirvani, 2020).
RF, when used in predictive models, is capable of identifying complex interactions between variables, especially when the response-predictor relationships are non-linear and change spatially (Stan and Sanchez-Azofeifa, 2019; Zanella et al., 2017). It also provides better spatial accuracy compared with other models (Prasad et al., 2006) and has proven to be less sensitive to the removal of variables in comparison with other algorithms (Hart et al., 2019). It is a powerful tool with the ability to estimate the relative importance of the predictive variables, and also reduces the risk of overfitting, which is that the model is good at predicting the data used for training, but performs worse with independent test data (Willcock et al., 2018). Additional features of the use of RF are that it can provide spatially explicit prediction probabilities, which can be extremely useful for decision-making (Saha et al., 2020).
Most of the studies that used RF as a predictive tool for deforestation risk found that biophysical and climate factors were more important than socioeconomic ones and that accessibility, distance to markets, and topography were always among the most influential factors driving deforestation. Although this statement can be questioned, since accessibility and distance to markets are human-induced factors, but they are controlled by biophysical ones (Aide et al., 2013; Cushman et al., 2017; Redo et al., 2012; Saha et al., 2020; Zanella et al., 2017).

Monitoring forest cover and environmental services throughout time using remote sensing: The topography factor

Remote sensing (RS) and Geographic Information Systems (GIS) techniques are essential tools that have increasingly been employed in Costa Rica and worldwide to monitor forest cover changes and the associated changes in environmental services provided by forests (De Araujo Barbosa et al., 2015; Sader and Joyce, 1988; Stan and Sanchez-Azofeifa, 2019; Vallet et al., 2016). Current trends in satellite imagery, its free availability, and wide access to large-scale cloud computing like the Google Earth Engine platform assure that the use of these techniques will be increasing (De Araujo Barbosa et al., 2015; Gorelick et al., 2017).
Vegetation indices (VI), defined as the arithmetic combination of two or more bands related to the spectral characteristics of vegetation (Liu and Huete, 1995; Rouse et al., 1973), have been used in a variety of fields including phenology, classification of vegetation, photosynthetic activity, aboveground net primary productivity and land surface temperature (Cao et al., 2016; Liu et al., 2020). Vegetation indices, particularly the Normalized Difference Vegetation Index (NDVI), are essential components of any study aiming to investigate environmental services especially those where vegetation, water, and biodiversity are involved (Cord et al., 2017; De Araujo Barbosa et al., 2015).
However, VI sensitivity is affected by the changing radiance that accompanies changes in orientation of the vegetation surface being sensed (Matsushita et al., 2007). The radiance changes at different times in the year and between years, due to different solar incidences over the surface, the so-called sun-sensor geometry (Teillet et al., 1982). Radiance is further changed in rough terrain, where a combination of the orientation of the terrain and the position of the satellite will determine high or low illumination conditions (IC).
The Enhanced Vegetation Index (EVI), in turn, is more sensitive than NDVI to biophysical attributes such as the Leaf Area Index (LAI) (Galvão et al., 2016; Peng et al., 2018) and much more affected by IC than NDVI, because is not a ratio-based VI and cannot compensate for variations in IC. In addition, EVI was proven to be five times more sensitive than NDVI to changes in Near-Infrared reflectance (NIR) (Galvão et al., 2016; Maeda et al., 2014; Maeda and Galvão, 2015; Peng et al., 2018).
Due to the effect of IC, the use of EVI as an indicator of vegetation functioning or forest productivity has been put under debate in recent years. Some authors claimed that the unusual greening effect observed in the dry season in the Amazon forest using EVI Moderate-resolution Imaging Spectroradiometer (MODIS) imagery (Huete et al., 2006) was induced by changes in sun-sensor geometry and not as a result of canopy structure, phenological patterns, or vegetation functioning (Morton et al., 2014). This effect has been confirmed in similar ecosystems at different times of the same season, with different IC (Galvão et al., 2011; Maeda and Galvão, 2015). However, in similar tropical forests, after removal of the IC effects, seasonal patterns were still present and seemed to be correlated with gross primary production (GPP), although authors recommended being cautious with this correlation (Maeda et al., 2014). Effects related to sun-sensor geometry and topography were also found at different times of the year using EVI and NDVI in subtropical deciduous forests under different IC. Sunlit and shadowed surfaces showed respectively different intensities of decrease and increase in reflectance even after topographic correction (Galvão et al., 2016).
In Costa Rica, the most valuable cloud forests in terms of biodiversity and PWAs are frequently located in low accessible areas, at high altitudes and irregular terrain (Bernard et al., 2009; Pfaff et al., 2009b). It is expected that topography will affect VI and consequently the evaluation of forest cover change or the environmental services derived from the calculation of these indices.

Research objectives and main research questions

This research aims to examine deforestation patterns in the cloud forests of Costa Rica and understand the biophysical and socioeconomic drivers behind them. Based on this analysis, develop a predictive model of deforestation risk that can be used to improve additionality and efficiency in the design of forest conservation schemes or policies such as the program of payments for environmental services. Additionally, it will examine the use of remote sensing techniques throughout time to evaluate their use as a tool to monitor forest cover and its environmental services associated.
In this sense, the four following specific objectives were established, namely:
1) To examine historical deforestation trends in the area and describe main biophysical and socioeconomic drivers
2) To use historical deforestation trends to develop a spatially explicit model of deforestation risk
3) To evaluate the effect of existing forest protection policies on historical and predicted deforestation trends and discuss their efficiency and additionality
4) To examine the use of remote sensing to monitor forest cover and the provided environmental services in the long-term
The main research questions to respond to these objectives is:
• How biophysical and socioeconomic factors influence historical and predicted deforestation?
• How the historical and predicted deforestation reveals the efficiency and additionality of PES?

Table of contents :

1. Introduction
1.1. Context, challenges, and stakes
1.2. Forest dynamics and conservation in Costa Rica
1.2.1. Protected wildlife areas and deforestation
1.2.2. Program of payment for environmental services and deforestation
1.3. Predictive models to evaluate forest risk loss
1.4. Monitoring forest cover and environmental services throughout time using remote sensing: The topography factor
2. Research objectives and main research questions
3. Methods
3.1. Study area
3.2. Biophysical approach
3.2.1. Protected wildlife areas
3.2.2. Private farms visited and delineated
3.2.2.1. Biophysical differences between PES and non-PES farms
3.2.3. Building a predictive model of vegetation loss risk
3.2.3.1. Selection of historical vegetation cover and vegetation loss data
3.2.3.2. Predictor variables used to train the model
3.2.3.3. Random Forests to model vegetation risk loss
3.2.3.4. Validation method
3.2.3.5. Variable Importance
3.2.4. Analysis of the topography and illumination condition (IC)
3.2.4.1. Landsat datasets and Image processing
3.2.4.2. Illumination Condition (IC)
3.2.4.3. Statistics across the collection of images
3.3. Socioeconomic approach
3.3.1. Semi-structured interviews and farm typology
4. Results
4.1. Vegetation risk loss
4.1.1. Random Forests to model vegetation risk loss
4.1.2. Validation of the model
4.1.3. Importance of the predictor variables
4.1.4. Historical and predicted vegetation loss risk between protected wildlife areas and unprotected areas
4.2. Farms dynamics
4.2.1. Typology of farms
4.2.2. Land use share, forest and opportunity cost
4.2.3. Historical and predicted forest loss risk inside delineated farms
4.2.4. Analysis of farms participating or not in the PES program and deforestation risk probability
4.3. Analysis of the topography and illumination condition
4.3.1. Illumination condition and vegetation indices
4.3.2. Time series for IC, EVI, and NDVI from 1984-2017
5. Discussion
5.1. A predictive model for vegetation loss risk: The importance of topography and accessibility
5.1.1. Accuracy and validation of the model
5.1.2. Importance of predictor variables
5.1.3. Analysis of historical vegetation loss and predicted vegetation/forest risk loss in PWA and farms
5.1.3.1. Historical vegetation loss in PWA
5.1.3.2. Predicted vegetation loss in PWA
5.1.3.3. Historical forest loss in farms
5.1.3.4. Predicted risk of forest loss in farms
5.1.3.5. Evaluation of deforestation risk probability in PES and non-PES forested areas
5.2. Far􀅵ers’ participatio􀅶 i􀅶 PES: La􀅶dscape i􀅶flue􀅶ces far􀅵i􀅶g strategies
5.2.1. Farm typology, opportunity cost, and proportion of forest area
5.2.2. Farmers’ perceptio􀅶 of forests a􀅶d the PES progra􀅵
5.3. The topography factor and its influence in the monitoring of forests and environmental services using remote sensing
5.3.1. Illumination conditions and vegetation indices
5.3.2. Temporal analysis of illumination conditions
6. Conclusions
6.1. Predictive models of vegetation loss risk in protected areas
6.2. The influence of topography on land use, opportunity cost, and participation in the PES program
6.3. The influence of topography in remote sensing and in the monitoring of forests and environmental services throughout the time
7. References

GET THE COMPLETE PROJECT