Time-Series Analysis of Deforestation and Land Conversion in Rondonia, Brazil: Insights for Vegetation Dynamics

Abstract

Monitoring the Brazilian Amazon has had a great deal of attention during the last decades as high rates of deforestation impact vegetation dynamics. The objective of this project was to use Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI)/Thermal Infrared Sensor (TIRS) data for the period 2003 to 2023 to compute a time-series analysis of deforestation and land conversion in Rondonia, Brazil. Land use and land cover (LULC) were examined over a 67,170 ha area of primary forest to derive patterns of photosynthetic change and plant water stress with respect to area of forest lost. A principal components analysis (PCA) was performed in the R programming environment to transform the Landsat scenes into six principal components (PC), and spectral indices including the normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI) were calculated. Using these data, a supervised classification with the Random Forest (RF) classifier was performed to classify each scene, and change was evaluated between them. The highest overall accuracy between classifications was 95.52% (kappa of 0.90). Finally, a time-series analysis was produced in Google Earth Engine (GEE) to compute 20-year median NDVI and NDWI trends for a collection of 195 Landsat scenes of the study area. A 59% loss of primary forest over the two decades analyzed is considered to be associated with a decrease in median NDWI by 0.018.

Introduction

Before 1975, deforestation in the Brazilian Amazon was less than 1%; however, the deforestation rate has increased exponentially in the last five decades, resulting in nearly 20% (776,996 km2) of the forested area being cleared as of 2021 (Moran, 1993; Ayad et al., 2024). This transition of natural forest ecosystems to pastureland and agricultural crops has a direct impact on vegetation dynamics, namely, related to photosynthetic activity and plant water stress. According to Uribe & Dukes (2021), differences in photosynthetic rates and water-use efficiency between the original and the replaced vegetation can affect the climate system on a regional and global scale. Despite numerous estimates of deforestation using satellite imagery, several uncertainties exist, including estimates on the rates and extent of photosynthetic change and plant water stress with respect to area of forest lost. Remote sensing (RS) of land cover change using spectral indices is a powerful means to monitor these changes, and especially time-series analysis has the potential to reveal long-term vegetation dynamics for forest management (Kuenzer et al., 2015). 

Rondonia, a western state in Brazil, continues to experience high loses of forest cover annually. From 2014 to 2016, Rondonia was responsible for approximately 21% of the total deforested areas in the Legal Amazon (Piontekowski et al., 2019). Therefore, because of the high deforestation rates in Rondonia relative to other Amazonian states, Rondonia is thought to represent patterns of photosynthetic change and plant water stress. The objective of this project was to use multi-date Landsat data for the period 2003 to 2023 to compute a time-series analysis of deforestation and land conversion. In particular, the following questions are addressed: 

1.  What are the long-term patterns of deforestation in the Brazilian Amazon using remote sensing-derived data?

2.  Is there a significant difference in photosynthetic rates and plant water stress with respect to area of forest lost?

Methods

A. Study Area

Land use and land cover (LULC) were examined from 2003 to 2023 over a 67,170 ha area of primary forest in Rondonia, Brazil. The study area is centered at approximately 63° 12' W and 8° 39' S and about 76 km northeast of the State’s capital, Porto Velho (Figure 1). It is adjacent to both the Jacunda National Forest and the Jamari National Forest, the latter of which contains the Santa Barbara tin mine. Rondonia is predominantly classified as hot, humid, and equatorial, with mean annual precipitation typically exceeding 2000 mm (Alvares et al., 2014). The study area was selected because of the recent deforestation activity at the onset of the analysis (since the early 2000s). In addition, the availability of relevant ground-based data of land cover in the area assisted the interpretation of LULC types (Lu et al., 2004). 

B. Data

To enable the analysis of deforestation and land cover change, Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI)/Thermal Infrared Sensor (TIRS) data were acquired for 2003 and 2023, respectively. The Landsat 5 TM scene was collected on July 15th, 2003, and the Landsat 8 OLI/TIRS scene was collected on July 22nd, 2023. The path and row of the scenes are 232 and 66, respectively, with 0% cloud cover. Selection of the scenes was based on criteria such as Collection 2 Level-2 and Tier 1 product availability, which are already corrected to surface reflectance and are therefore suitable for time-series analysis (USGS, 2021). Hence, radiometric correction was not required for these data. False-color composites using the shortwave infrared 1 (SWIR1) band in the red channel, the near-infrared (NIR) band in the green channel, and the red band in the blue channel were produced for each original scene to aid interpretation of vegetation change between time stamps (Figure 2). Both scenes were cropped to a 67,170 ha extent where significant forest conversion was thought to occur during the 20-year period. 

C. Change Detection

Two-date image differencing techniques are long-established methods for detecting LULC change that produce easily interpretable results (Singh, 1989). Following image selection and preprocessing (when necessary), the process ideally involves (1) data transformation, such as calculating spectral indices of interest; (2) classifying the images using threshold or supervised classification techniques; and (3) evaluation. Accordingly, a principal components analysis (PCA) was performed in the R programming environment to transform each scene into six principal components (PC). The last three PCs of each scene were discarded due to their very low variance and limited information content (Figure 3). In addition, spectral indices including the normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI) were calculated using the equations (NIR - Red)/(NIR + Red) and (NIR - SWIR1)/(NIR + SWIR1), respectively (Figure 4). NDVI is used to quantify vegetation greenness (photosynthetic activity) and is useful in deriving changes in plant health; NDWI is used to quantify vegetation moisture and is useful in deriving changes in plant water content (USGS, 2018; Gao, 1996). By only including the first three PCs, the NDVI, and the NDWI, a higher classification accuracy was anticipated. 

The Random Forest (RF) classifier was selected for supervised classification because of its ability to produce LULC maps with relatively high accuracies through open-source environments (Ristin et al., 2016). Although ground data were not acquired, availability of relevant ground-based data of land cover in the area and high-resolution satellite imagery in the Google Earth platform assisted visual interpretation of LULC classes. Moreover, Lu et al. (2004) identified seven land cover classes in northeastern Rondonia from fieldwork conducted during the dry seasons (June–August) of 1999 and 2000: water, bare lands, agricultural lands, pasture lands, initial secondary succession, advanced secondary succession, and mature forest (p. 727). Considering the relevance of their work, and using high-resolution satellite imagery as a reference, four simplified land cover classes were visually interpreted in the study area: open water, pasture/agriculture, primary forest, and secondary forest.

Training data for supervised classification were created in ArcGIS Pro with the Training Samples Manager and subsequently partitioned into training and validation data in R. The RF classifier was used to classify and validate each scene with the extracted values from the first three PCs, the NDVI, and the NDWI, and change was evaluated between scenes on a pixel-by-pixel basis. Using a standard conversion to transform pixel size to area (30 x 30 m = 900 m2 = 0.9 ha), total area of change for each class was derived. Finally, in conjunction with the classified Landsat 5 TM and Landsat 8 OLI/TIRS maps, a time-series analysis was produced in Google Earth Engine (GEE) to compute NDVI and NDWI trends for a collection of 195 Landsat scenes cropped to the 67,170 ha study area. Cloud cover was filtered to <10% and the date range was set to the 20-year study period. As a result, differences in photosynthetic rates and plant water stress with respect to area of forest lost were inferred.

Results

A. Accuracy Assessment

The confusion matrices resulting from the RF classifications of the 2003 and 2023 scenes are presented in Table 1 and Table 2, respectively. The classification of the 2003 scene had the highest overall accuracy of 95.52% (kappa of 0.90). The classification of the 2023 scene had a lower but acceptable overall accuracy of 77.06% (kappa of 0.69). Because open water exhibited similar brightness to areas of pasture/agriculture, it was difficult to discriminate between them, causing errors of omission and commission. Therefore, some of the errors may be overestimated due to the expansion of heterogenous pasture/agriculture in the study area over time.

B. Change Analysis

Clearing of primary forest was evident and easily visualized in the final classified maps (Figure 5). As of 2023, approximately 21,368 ha, or nearly 32% of the original 66,423 ha of primary forest in 2003, had been cleared for pasture/agriculture. About 17,539 ha of additional primary forest were in stages of regrowth (secondary forest), and 40 ha were lost to open water. Only 58 ha of primary forest were gained (less than 1% of the study area) during the two decades analyzed. In total, approximately 38,889 ha, or nearly 59% of the original 66,423 ha of primary forest in 2003, had been cleared or lost (including areas in regrowth) by 2023 (Table 3). 

C. Time-Series Analysis

The NDVI and NDWI trends resulting from the time-series analysis in GEE are presented in Figure 6 and Figure 7, respectively. Most long-term RS time-series analyses consist of three components: (1) a long-term directional trend, or the line of best fit for all observations; (2) seasonal, systematic movements; and (3) irregular, unsystematic short-term movements (Kuenzer et al., 2015). Here, the focus was on long-term directional trends of NDVI and NDWI, which were to be analyzed with respect to their 20-year median trend. For NDVI, an ascending trend component was present despite seasonal fluctuations likely indicative of shifting cultivation and regrowth. A descending trend component was present for NDWI while also exhibiting seasonal fluctuations. The 20-year median NDVI trend ascended from about 0.364 to 0.367, whereas the 20-year median NDWI trend descended from about 0.194 to 0.176. Both trends used the ee.Reducer.median() function for higher accuracy.

Discussion

Previous studies in Rondonia have indicated that brightness and greenness will generally be lower for mature or primary forests than for regrowing vegetation due to shadowing from variation in canopy heights in later stages of succession (Guild et al., 2004). Therefore, given the results in this study, it is inferred that long-term median NDVI trends do not directly correspond to photosynthetic activity by primary forest in stages of regrowth. Rather, in cleared areas undergoing regrowth, greenness will generally increase to the point of canopy closure, hence the lower median NDVI values at the onset of the analysis. The amount and condition of photosynthetic activity of vegetation greenness can be captured through RS measurements (Haboudane et al., 2008), but due to the complexity of shadowing, accurate estimates of differences in photosynthetic rates over time are difficult to evaluate using RS data alone. In addition, Schwartz et al. (2019) found that loss of canopy water and decline in photosynthesis are not necessarily coupled at the landscape scale (p. 9), which may be indicated by the conflicting trends in this study. Long-term median NDWI may be an effective estimate of leaf water content irrespective of shadowing. Therefore, a 59% loss of primary forest over the two decades analyzed in the study region is considered to be associated with a decrease in median NDWI by 0.018. Further research might include the comparison of different spectral indices. 

Conclusion

The importance of this project was the reliable detection of deforestation patterns in the Brazilian Amazon using remote sensing-derived data and the inferred long-term median NDWI trend with respect to area of forest lost. Although differences in photosynthetic rates could not be evaluated, this project demonstrates that RS of land cover change using spectral indices is a powerful means to monitor deforestation, and that time-series analysis can reveal long-term vegetation dynamics for forest management. Finally, the results in this study may provide interesting data regarding the impact that deforestation is having on the water cycle in Amazonia.