3.1 Case study

Among the countries of the Congo River Basin, which include Cameroon, the Central African Republic, the Republic of Congo, Equatorial Guinea and Gabon, the Democratic Republic of Congo retains the highest share of dense forest, for a total of approximately 155.500.000 ha [27, 28]. Forest accounts for about 55% to 67% of the national territory, depending on the definition of forest adopted. The typologies of forest vary throughout the country, depending on the different agro-climatic zones that the country encompasses. The most common forest typologies are: dense rainforest, swampy forest, dry forest and forest–savannah mosaic [28]. The presence of these various biomes is conditioned by the high diversity of climates in the country. Mean annual precipitations fluctuate around 1,500 mm per year but vary consistently by zone. For example, along the west coast annual cumulative rainfall is only 800 mm, while near the Equator it reaches 2,000 mm. The spatial variability is coupled with a strong inter-annual variability: the nethermost areas are characterized by two rainy seasons (March to May and September to December), while the southeastern area has a single rainy season (July to August) [29]. As shown in the data section, these seasons are reflected in the agricultural seasonality cycle [30].

3.2 Deforestation in DRC

Most studies addressing the deforestation dynamics in the DRC are generally part of a wider analysis which focuses on the entire perimeter of the Congo River basin. Among others, the report on deforestation trends by the World Bank [27] suggests the importance of an exclusive focus on the case of the DRC: the report highlights that while deforestation trends in the tropical forests of the DRC are still lower than those observed in neighboring countries, forest degradation is increasing. This may be due to the peculiarities of the country, where sociopolitical instability has prevented the installation of a widespread agriculture and logging industry. In fact, despite being the country with the largest share of biomass and forested area, the DRC has the lowest level of logging activity [27] in in the region. Indeed Fuller et al. [31] assessed the impacts of China’s wood commodity imports from the Congo basin finding the trade from DRC to be negligible with respect to trade with its neighboring countries. This result highlights the fact that forest loss is driven primarily by slash-and-burn agriculture rather than by the commercial timber harvest. Similarly, Tyukavina et al. [32] found small-scale clearing the most dominant form of deforestation: small households were accountable for more than 90% of the total DRC forest cover loss.

Unfortunately, despite the DRC’s peculiarities in terms of deforestation drivers and rates, which make the country an interesting case study, the literature addressing the DRC’s deforestation factors is rather limited. Achille et al. [33] investigated the DRC’s dense forest clearing and degradation dynamics between 1990 and 2018. The authors showed that the total net deforestation rate was around 2.12%, while degradation was 0.12%, with a substantial acceleration after 2005. The authors identified four main deforestation drivers in particular: 1) the population’s heavy dependence on wood for fuel; 2) the practice of itinerant agriculture 3) the presence of mining quarries (also addressed by Davis et al. [34] and Butsic et al. [35]); 4) poor regulatory capabilities in the enforcement of land use rights and land protection. Results are in line with Molinario et al. [36], who investigated the period between 2000 and 2015. The drivers identified by Achille et al. [33] are also potentially connected to the sociopolitical instability, which is recognized as having significant effects on deforestation in developing countries [37]. For example, Butsic et al. [35] found that conflicts in the DRC are a significant driver of deforestation. The authors also found protected areas effective in preventing land clearing, even in the presence of conflicts.

In the DRC, deforestation and degradation are mainly driven by small-scale, subsistence and mostly rainfed agriculture. Given the significant weather variability in the country, addressing how weather shocks, adaptation strategies and forests are connected is of crucial importance, especially considering the fundamental insurance role played by forest biomes in the Congo basin [38]. As mentioned before, the studies addressing the links between deforestation and weather shocks are even more limited, and none have focused on the DRC. For example, while studying the effect of the poor rainy season in Western and Central Africa, Leblois [25] also considers a small portion of the north-western DRC’s tropical forests. Similarly, other global studies, such as Zaveri et al. [22], address the issue of rainfall anomalies in developing countries but do not examine local dynamics and socio-economic factors.

3.3 Data description

To investigate the effects of droughts on deforestation, we use three main georeferenced data sets. Final data, information on how these were retrieved and processed, as well as codes for the econometric analysis are available in the data repository supporting this publication [39].

First, the Global Forest Change 2000–2020 database by Hansen et al. [40] offers the fundamental information used to define our dependent variable:

1. a. the percentage (0–100%) of tree canopy cover for the year 2000 per output grid cell, which represents the canopy closure for all vegetation greater than 5m in height. This variable was provided as a georeferenced database composed by 1 arc second gridded pixel cells (i.e., measuring approximately 30 by 30 meters at the Equator). It was then reshaped to meet the definition of the precipitation database (0.05 decimal degree, around 30km² at the Equator, set as a standard for our unit of the present analysis).
2. b. a mask layer indicating the state of a cell between mapped pixels, waters or unmapped areas. The mask layer was also reshaped to obtain the total mapped ha for each cell.
3. c. the year of gross forest cover loss during the period 2000–2020, built as a dummy for each cell indicating whether it was deforested and in which year. A cell is considered as deforested by Hansen et al. [40] only when a clear forest cover cut is observed, i.e. the cell is characterized by a ∼0% crown cover replacement. This selection isolates forest loss attributable to human activities while excluding canopy disturbances.
   Through the aggregation procedure to meet the standard resolution set for the present analysis, this variable was reshaped and transformed to obtain the percentage of each cell deforested every year. We then calculated the total forested area lost for each year in each pixel, through the combination of the data on forest cover loss and the mask layer on mapped areas (point b). The lost area, expressed in hectares, correspond to the final dependent variable applied in the analysis.

Forest cover in the year 2000 and the average yearly loss, by pixel, are represented in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Forest cover in the year 2000, average deforestation 2001–2020, by cell and by agricultural region.

Source: GADM [49], Hansen/UMD/Google/USGS/NASA [40] and Crop Calendar Dataset [30].

https://doi.org/10.1371/journal.pone.0276667.g002

A second essential data set is the Crop Calendar proposed by Sacks et al. [30], which provides the georeferenced agricultural calendar for multiple crops. It allows us to divide the country into homogeneous areas of production based on the crop’s agricultural cycle and compute the duration of the planting, growing and harvesting season for each cell and crop.

Due to the importance of maize and cassava in the local diet (respectively 50% and 33% of the total daily intake), [41] these two crops were selected to define the cropping-homogeneous regions of the DRC. Eventually we distinguish four regions (north, center, south and extreme south), each of which is characterized by a different cropping calendar (see data paper).

The cassava planting season extends from April to August in the north, and from October to February in the center, south and extreme south regions and has a single cropping cycle per year. Maize, on the other hand, is characterized by a double cropping cycle per year, so that in the north the planting seasons are from February to March and from June to July. A notable exception is the extreme south region (corresponding to the Haut-Katanga province), which is characterized by a single cropping season even for maize. For this reason, this region was excluded from the analysis.

The duration of seasons in each region was used to estimate a seasonal drought index, based on the average Standardized Precipitation Index (SPI) [42]. The SPI is a standardized index of precipitation quality which informs how much a rainfall observation at a specific time deviates from its long term mean. The index assumes the value of zero whenever the observed rainfalls are around the mean, a positive value (up to +3) for positive deviations, (i.e., a surplus in rainfall) and a negative value (up to -3) when there is a deficit in rainfall. Commonly, the SPI threshold to consider a deficit as a drought is -1. The SPI was calculated on the basis of the monthly precipitation (retrieved from CHIRPS [43]). The drought intensity associated with each grid cell was thus obtained by averaging the monthly SPI for each crop season. Finally, a synthetic drought index was developed, taking value equal to 1 whenever the seasonal average SPI was falling below minus one, and 0 otherwise (dummy variable). A similar approach to seasonality was undertaken by Noack et al. [1]. In particular, the authors used the agricultural calendar to quantify the shares of households’ income generated in each season and thereafter the effect of seasonal droughts on income. Nevertheless, the drought index the authors applied is computed as a continuous variable expressing water endowments anomalies associated to the income periods.

Two supplementary variables were added to investigate potential heterogeneity sources. To assess the effectiveness of public policies that addressed forest protection, we used the World Database on Protected Areas (WDPA by UNEP-WCMC [44]). The database provides georeferenced polygons delimiting any protected areas, as well as information about their status, including the year in which the protection was put in place. We then associated a fixed dummy variable to each cell to indicate whether or not the observation was included in a protected area in the year 2000. Lastly, the proximity to cities was retrieved from the Travel Time to Major Cities database [45], which provides for each cell the estimated time required in the year 2000 to reach the nearest city of at least 50,000 inhabitants. Lastly, a dummy variable was created to control for proximity indicating whether a cell was closer than four hours of travelling from a city.

From the processing and merging of the above-mentioned data, we obtained a disaggregated panel database that was constituted of grid cells of approximately 30km² at the Equator. Due to the peculiarity of the original Global Forest Change database, the present analysis covers the years from 2001 through 2020. We use as a benchmark the original state of the forest and the enforced protected areas in the year 2000 (pre-analysis levels) in order to prevent our assessment from potential reverse causality. The final panel database was filtered to maintain only cells that had on average a 50% forest cover in the year 2000: the choice, which allows to retain medium to high forest cover areas, is based on similar thresholds applied by Achard et al. [46, 47] and Verhegghen et al. [48]. For computation matters, the panel was then further randomly sampled to maintain only 35% of the observations, corresponding to 25,958 pixels observed over 20 years (2001–2020).

3.4 Descriptive statistics

In the year 2000, the average forest cover of each observation was 86%. Our analysis focuses on forested areas: as highlighted in the data section (3.3), we consider a cell as forested if it had at least 50% of cover in the year 2000. Between 2001 and 2020, in each of these cells, an average of 13 hectares was lost every year (about 0.43% of their area). This result is consistent with results obtained by multiple authors. Achille et al. [33] found a yearly deforestation rate of 0.42% between 1990 and 2018. Our indicator is slightly higher than the average deforestation rate observed by other authors which instead spans from 0.15% to 0.4% in function of the years considered (results resumed by Kengoum et al. [28]). This gap should not surprise, as it can be explained by the increasing trend already observed by Achille et al. [33] and suggested by the World Bank [27] or simply by the particular definition of forests used by each author.

The average cumulative annual rainfall in the DRC reaches 1,692 mm. However, the very large variability in the descriptive statistics (S1 Table in S1 Appendix) can be explained by the country’s wide extension, which defines both the spatial heterogeneity in climate and the observed seasonal intra-annual variations over crop-specific agricultural calendars. For instance, the cumulative rainfall over the cassava agricultural cycle (one per year) ranges from 849 mm during the planting season to 844 mm during the harvesting season. For what concerns the maize’s two annual agricultural cycles, the average cumulative rainfall during the planting seasons is 319 mm (with the first season receiving an average of 100 mm of rain more than than the second). During the growing and harvesting seasons, precipitation is approximately 740 mm and 400 mm, respectively.

The observed cells, which covers the entire country with the exclusion of the Haut-Katanga province, are located at an average travel distance of nine hours and 43 minutes from the nearest main city.

3.5 Identification strategy

As mentioned earlier, we distinguish 1/ current and experienced droughts; and 2/ the timing of the drought, to assess how they relate to deforestation.

On the one hand, responses to current droughts are expected to have short-term effects related to coping strategies that may be implemented once the shock occurs. On the other hand, the experience of one or multiple droughts in past years is expected to induce long-term effects related to the outcomes of adaptation strategies.

Our dependent variable L_it is the number of hectares of forest lost in cell i at year t. The explanatory variables are dummies indicating whether or not there has been a drought during any of the cassava or maize agricultural seasons. A season is considered to be affected by a drought whenever the average monthly SPI falls below -1: in that eventuality, the explanatory variable assumes value equal to 1 and zero otherwise. An important factor is that while the cassava agricultural calendar is composed of a single cycle and only two seasons (planting and harvesting), maize is characterized by two cycles per year (planting/growing/harvesting). Thus, two planting, two growing and two harvesting seasons are observed each year.

In our analysis, both maize cycles are considered jointly. The dummies take the value 1 independently from the cycle that has been hit by drought; that is, if in year t a drought hits the cell i only in the maize planting season of the first agricultural cycle, then variable Droughts_Maize_PL1, 2_it takes the value of 1. Similarly, if both the first and second planting seasons are hit by drought, then variable Droughts_Maize_PL1, 2_it also takes the value of 1.

In order to estimate these impacts, a fixed-effect panel linear regression model is applied. Short-term effects and seasonality are investigated through the following Eq (1): (1)

The same methodology applies when studying the effects of experienced droughts on current deforestation (Eq (2)). The effect of experienced droughts, as in the aim of model (2), is considered as the effect of any drought that occurred in the 3 years prior to the observed deforestation. Again, the dummy is equal to 1 whenever at least one drought has occurred during any of the two agricultural cycles in the three previous years. (2)

In both the models (1) and (2), γ_i, δ_t, and u_it represent the cell fixed effect, the time fixed effect and the error term, respectively. Standard errors are clustered at the sector level [50], corresponding circa to level 3 of the Global Administrative Area Database (which includes provinces, districts, territories and sectors) [49].

In S2 Appendix, the same models are further developed to better investigate the effect of droughts in each of the two agricultural cycle of maize. Thus, we developed an explanatory variables, again constructed as a dummy, for each of the six maize agricultural seasons observed over a year (first cycle plant, growing, harvesting and second cycle plant, growing, harvesting).

The model of Eq (1) is further developed to investigate the sources of heterogeneity. In particular, we investigate:

1. a. the role of protected areas. The relationship was estimated by interacting the explanatory variable with a dummy indicating the presence of a PA. More precisely, whenever deforestation was taking place in an area that was protected in the year 2000 (corresponding to the year prior to the beginning of our analysis);
2. b. the role played by the proximity to cities. The relationship was estimated by interacting the explanatory variable with a dummy indicating whether the deforested cell is at a travel distance of less than four hours from a main city (i.e., a city is considered to be of interest when registering more than 50,000 inhabitants in the year 2000);
3. c. the role played by the baseline forest cover. Two approaches are undertaken: First by repeating the models presented in Eqs (1) and (2) but maintaining respectively only observations characterized by low forest cover (i.e., minor or equal than 75% in year 2000) or by high forest cover (i.e., higher than 75% in year 2000); Second by interacting both the models with a dummy indicating observations with high forest cover (i.e., assuming value equal to 1 whenever the baseline forest cover is higher than 75%).

Lastly three main robustness checks are undertaken by:

1. a. converting the dependent variable using an inverse hyperbolic sine transformation of the forest loss, that allows us to retain the multiple zeros present in our database (Summary statistics available in the S1 Appendix). Therefore, the dependent variable of both (1) and (2) is substituted by arcsinh(L_it) while the explanatory variables remain unchanged;
2. b. reducing models presented in Eqs (1) and (2) to individually study cassava and maize cycles;
3. c. excluding from the panel years affected by the Second Congo War.

4.1 Seasonality matters

The main purpose of this paper is to investigate the effects of droughts on deforestation; first, discerning between experienced and current droughts, and second, investigating the role of seasonality. Again, by seasonality, we mean a period of the agricultural cycle—either planting, growing or harvesting. Recalling that the cassava agricultural cycle is composed of only two seasons, planting and harvesting, is of upmost importance. This peculiarity is due to the reproduction methodology: as cassava is diffused by stems obtained by the plants cropped in the prior agricultural season, and not by seeds, the growing period is less observable with respect to other crops where germination time is required.

4.1.1 Experienced droughts.

Experienced droughts over the entire agricultural period do not appear to influence deforestation (S2 Table, column in S2 Appendix). Yet distinguishing the seasons allows us to sharpen the analysis. Table 1, column (1) illustrates the effects of experienced droughts, i.e., those that occurred during the three years before the observed deforestation. While previous shocks during the cassava agricultural cycle do not seem to induce any deforestation, experienced droughts hitting the maize growing or harvesting season tend to decrease deforestation in the following years. More precisely, droughts that occurred during the maize growing season reduce current deforestation by 2.6%, i.e., of 0.33 hectares less than the average 12.96 hectares lost on average in each cell; while they reduce it during the harvesting season by 1.9%. The overall consequence of experienced droughts is reduced deforestation. This result suggests that household adaptation strategies may imply a reallocation of the activity portfolio from agriculture to off-farm activities, including a higher reliance on NTFP harvesting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Deforestation in response to experienced and current droughts.

Analysis of cells characterized by at least 50% forest cover in the year 2000. The observations are then sampled to maintain the 35% of their total.

https://doi.org/10.1371/journal.pone.0276667.t001

These results are in line with the results from Desbureau & Damania [20], who found that repeated droughts in the past tend to decrease deforestation. This outcome was also predicted in a theoretical model by Delacote [7], who suggest that a higher perceived risk in agriculture may lead households to diversify their livelihood strategy, reducing their share of agriculture but relying more on forests, which would lead to a lower deforestation rate. Nevertheless, a household’s higher reliance on NTFP may imply a higher level of forest degradation [8], which has not been investigated in this paper.

Interestingly, when we exclude the war years prior to 2004 to check the robustness of the previously discussed results, results concern experienced droughts that occurred anytime over the agricultural year proved to be significant, confirming that experienced droughts tend to reduce deforestation. These outcomes suggest that there is room for further research investigating the interactions between conflicts, droughts and deforestation.

4.2 Looking for sources of heterogeneity in the land-use response to weather shocks

In this section, we investigate the sources of heterogeneity influencing deforestation, as a response to current droughts. The interactions with the first model are not presented, due to the already low significance of the effect of experienced droughts (Table 1, column 1).

In particular, we analyze the influence of conservation policies, economic pressure and forest cover. Conservation policies are represented by protected areas, while economic pressure is measured by proximity to cities, which entails lower transport costs and better outside options for households. The current drought indicators were interacted with these two variables. The main results are presented in Table 2 while the most significant ones are displayed in Figs 3 and 4. The influence of the forest endowment is instead investigated in function of the observed forest cover share in year 2000 (Tables 3 and 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Deforestation and current droughts; interaction with protected areas (installed before the year 2000) and proximity to main cities.

The analysis considers cells characterized by least 50% forest cover in the year 2000. Total observations are then sampled to maintain only the 35% of the total.

https://doi.org/10.1371/journal.pone.0276667.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Deforestation and droughts over the main agricultural periods and cycles, differentiating between low (≤ 75%) and high (> 75%) forest cover in year 2000.

https://doi.org/10.1371/journal.pone.0276667.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Deforestation and droughts over the main agricultural periods and cycles, interaction with high baseline forest cover.

Total observations are then sampled to maintain only the 35% of the total observations.

https://doi.org/10.1371/journal.pone.0276667.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Deforestation and current droughts; graphical representation of deforestation in response to experienced and current droughts.

https://doi.org/10.1371/journal.pone.0276667.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Deforestation and current droughts; graphical representation of interaction with protected areas and proximity to main cities.

https://doi.org/10.1371/journal.pone.0276667.g004

The general seasonality effects that are found in the previous section are robust to the addition of these interactions. Overall, protected areas tend to mitigate some (positive or negative) effects of droughts on deforestation, while proximity to cities tends to exacerbate them.

4.3 Robustness checks

We provide three main robustness checks of the previous models (displayed in S3 Appendix). In the first check, the dependent variable is transformed using an inverse hyperbolic sine transformation. Results, displayed in S3 Table of S3 Appendix are not significantly different than those obtained applying the model of Eq (1). Secondly,the cassava and maize cycles are individually studied. Again, intensity, direction and significance did not remarkably vary from the main models discussed in the results section (S4 and S5 Tables in S3 Appendix). Lastly, the DRC has been shaken by a long period of instability. The Second Congo War took place between 1998 and 2003. It affected the role of the State, compromising its functioning [35, 54], as well as households’ behaviors concerning land use and forest access. For this reason, the Eq (1) is estimated excluding the years of war. Again the significant variables maintain the same intensity and direction as in the previously discussed models (S6 Table in S3 Appendix).