Connecting the Dots
Summary
-
Satellite data from 1990 to 2024 show that intact forest with high NDVI is concentrated in remote areas, while zones near roads, cities, and rivers exhibit fragmented vegetation, likely due to agriculture, logging, and settlement expansion. Increased SWIR reflectance in these regions suggests soil exposure and surface drying from land clearing.
-
A moderate negative correlation (R² = 0.43) between NDVI and SWIR confirms that vegetation gain tends to reduce surface reflectance, while vegetation loss increases it. This dual-index approach helps distinguish between degradation and recovery across the landscape more effectively than NDVI alone.
-
Zonal statistics revealed that 137 km² of the study area showed vegetation gain, while only 4.87 km² showed loss, suggesting localized but ecologically important transformations, often along the forest frontier. Most of the landscape (96%) remained stable.
-
The NDVI–SWIR matrix identified patterns, such as degraded vegetation masked by stable NDVI or cropland regrowth with high moisture signals. Classes like 31 (true regrowth) and 13 (vegetation loss + drying) were especially important for identifying ecological transitions.
-
Ground photos and PlanetScope imagery provided essential context, confirming trends at sites like the water buffalo farm and plantation. However, mismatches in some locations highlight the challenges of interpreting single-date RENDVI in complex or post-disturbance environments.
-
Key constraints included differences in spatial resolution between Landsat and PlanetScope, single-date RENDVI data, and limited field access due to high-water conditions. These affected cross-validation and the ability to generalize findings across the entire region.
Landscape-Scale Patterns in NDVI and SWIR
At a large scale, the NDVI composite images from 1990 to 2024 reveal distinct spatial patterns of vegetation health and land use change along the Amazon River. High NDVI values correspond to intact forest areas with dense canopy cover, particularly in remote or less-accessible zones. These regions have likely remained undisturbed due to limited road access or conservation protections. In contrast, areas near major river channels, roads, and urban zones display more fragmented NDVI signatures indicating vegetation stress, clearing, or land conversion, likely caused by settlement expansion, fire, agriculture, or logging.These patterns are supported by the SWIR change detection image, which highlights increases in shortwave infrared reflectance, typically associated with exposed soils, urban infrastructure, or deforested land. Areas of increased reflectance in the SWIR difference strongly suggest expansion of impervious surfaces or cleared ground. These are predominantly seen around urban edges and along transportation corridors.
Linking Vegetation Trends with Surface Conditions
The scatter plot comparing NDVI and SWIR changes further supports the patterns observed through image analysis and zonal statistics. A moderate negative correlation (R² = 0.43) indicates that areas with increasing NDVI values tended to show a corresponding decrease in SWIR reflectance, a relationship consistent with forest regrowth and the restoration of canopy cover. Conversely, areas with declining NDVI often exhibited increased SWIR reflectance, highlighting vegetation loss and the exposure of dry, reflective surfaces such as bare soil or urban infrastructure. This relationship reinforces the conclusion that land use changes in the region are spatially heterogeneous, while some areas are undergoing ecological recovery, others are experiencing degradation. The scatter plot serves as a valuable diagnostic tool, quantifying the inverse relationship between vegetation health and surface exposure, and providing statistical support for the dual processes of regrowth and fragmentation identified across the study area.
Quantifying Vegetation Change
This interpretation is quantified using the Tabulate Area tool in ArcGIS Pro, which calculated vegetation change from a reclassified NDVI difference raster. Results showed that approximately 274.16 km² of the study area experienced vegetation loss (NDVI decrease), while 773.15 km² showed signs of vegetation gain (NDVI increase). The overwhelming majority, 32,659.41 km², or 96.89% of the landscape, exhibited no significant change. These proportions reflect a landscape where land use pressures have led to localized, yet ecologically significant, transformations, particularly at the forest frontier. It is also possible that some of the observed vegetation increase is due to hydrological changes in the river system, such as declining or shifting water levels that exposed river margins or floodplains, allowing vegetation to establish where open water previously existed.
The histograms provide a frequency distribution of pixel values for NDVI and SWIR changes between 1990 and 2024, offering a quantitative overview of vegetation and surface condition shifts. The NDVI histogram shows a strong concentration around zero, indicating that much of the landscape maintained relatively stable vegetation cover over time. However, the noticeable left skew, representing negative NDVI values, suggests substantial localized vegetation loss in some areas. The SWIR histogram similarly clusters near zero, but with a broader spread toward both positive and negative values, reflecting heterogeneous surface moisture and reflectance changes. Positive SWIR values imply increased soil exposure or drying, while negative values suggest moisture gain or canopy thickening. Together, these histograms support the matrix classification results, showing that although most of the study region remained stable, distinct areas experienced degradation, recovery, or transitional shifts in land use and ecological function.
Enhanced Interpretation Using NDVI–SWIR Matrix
The combined NDVI–SWIR classification matrix offers a more detailed understanding of landscape change than NDVI alone by integrating vegetation greenness with surface moisture trends. This dual-index approach allows for improved interpretation of ecological processes such as forest degradation, regrowth, and land conversion. Specifically, areas classified as NDVI loss with increased SWIR reflectance (Class 13) indicate likely vegetation removal with concurrent surface drying, characteristic of deforestation, urban expansion, or the clearing of vegetation that leaves soils more exposed to sunlight and evaporation. In contrast, NDVI gain paired with decreased SWIR reflectance (Class 31) suggests true vegetative regrowth, associated with increased canopy closure and improved soil moisture retention, consistent with natural forest recovery or successful reforestation.
Other class combinations reveal more subtle ecological dynamics. For example, areas exhibiting stable NDVI with increased SWIR reflectance (Class 23) may indicate degradation not visible in greenness indices alone, such as selective logging, understory clearance, or drying of shallow-rooted vegetation. Similarly, vegetation gain accompanied by increased SWIR reflectance (Class 33) may signal land conversion to crops or grasses, particularly irrigated or moisture-intensive monocultures, rather than forest regrowth. These interpretations underscore the value of incorporating SWIR data into landscape change analyses, as NDVI alone can mask degradation and misclassify land transitions with similar greenness but differing surface conditions.
The spatial distribution of matrix classes across the study area was further interpreted using PlanetScope 8-band high-resolution imagery and geotagged field photographs collected in June 2025. Several field sites provided critical ground context for validating classification outputs and understanding site-level variation. For instance, locations such as the water buffalo farm, enchanted lake site, and the passionfruit vine field exhibited visually stable vegetation cover and were classified under Class 21 (stable NDVI with decreasing SWIR), indicative of surface moistening. These observations may reflect persistent vegetation or land management practices that maintain vegetation while enhancing surface moisture. However, at the enchanted lake and passionfruit sites, RENDVI values were lower than expected, possibly due to the spectral influence of surface types or crop structure. As a result, these two sites showed only partial agreement with Landsat-based classifications.
In contrast, sites such as the new plantation and the burnt manioc clearing both showed low NDVI values, consistent with vegetation removal, but diverged in SWIR response. The burnt manioc site exhibited decreased SWIR, likely indicating early successional regrowth or organic residue retention. However, PlanetScope RENDVI showed a high vegetation signal that contradicted the low NDVI, resulting in a classification mismatch. The plantation site, on the other hand, showed reduced SWIR reflectance potentially linked to irrigation, mulch cover, or early-stage vegetation establishment and demonstrated better alignment with the RENDVI signal. These were classified as Class 21 and Class 23, respectively, but only the latter aligned with PlanetScope-derived vegetation measures, highlighting the need for caution when interpreting single-date RENDVI in post-disturbance environments.
Additional contrasts were observed in sites with vegetation gain. The cleared-land regrowth plot (Figure 9) and the plantation site both demonstrated increased NDVI yet differed in SWIR response. The regrowth plot had decreased SWIR, suggesting denser canopy recovery and increased moisture (Class 31), and closely aligned with RENDVI estimates of vegetation gain. Conversely, the plantation site had increased SWIR, possibly reflecting cropland expansion with shallow-rooted vegetation or exposed soil between rows (Class 33), but RENDVI values were unexpectedly low. This mismatch suggests that vegetation structure or sparse canopy may reduce the effectiveness of RENDVI in detecting crop-based regrowth. A third site, the resort fire scar, showed NDVI gain and SWIR loss, interpreted as genuine post-disturbance forest recovery and improved moisture retention, which matched well with moderate RENDVI values and Class 31 classification.
RENDVI: Strengths, Weaknesses, and Use Cases
The Red Edge Normalized Difference Vegetation Index (RENDVI) was a key metric used to assess current canopy health across the Amazon study sites. Unlike traditional NDVI, which uses red reflectance, RENDVI substitutes the red edge band, making it more sensitive to chlorophyll concentration and subtle stress responses in vegetation. This enhanced sensitivity allows RENDVI to capture fine-scale differences in canopy condition, especially in areas where regrowth is sparse, early-stage, or under environmental stress.
In this analysis, RENDVI was calculated using 2024 PlanetScope imagery to represent the most recent state of vegetation vigor at each field site. Because it focuses on current spectral reflectance rather than change over time, RENDVI was particularly useful for evaluating the quality of regrowth in areas where other indices, like NDVI and SWIR, produced mixed signals. For example, several sites (e.g., Plantation, Regrowth, and Passionfruit Vines) exhibited NDVI gain and SWIR increase, which could suggest either recovering vegetation or sparse regrowth under drier conditions. RENDVI values in these areas were negative or near zero, pointing instead to low canopy density or ongoing degradation, aligning more closely with field observations of limited recovery.
However, RENDVI also showed partial disagreement in some cases, such as Burnt Manioc and New Plantation, where its low values conflicted with moderate EVI or NDVI gains. These discrepancies likely reflect RENDVI's conservative nature and its focus on chlorophyll response, which may not yet reflect early or patchy regrowth visible in other indices. Additionally, RENDVI's performance was impacted by limitations in the input data, specifically, the use of unscaled digital numbers from PlanetScope imagery and visible striping from the mosaicked scenes, which introduced minor inconsistencies across the landscape.
EVI
The Enhanced Vegetation Index (EVI) is designed to measure vegetation density and canopy condition while minimizing atmospheric distortion, soil background noise, and saturation in areas of dense vegetation. Unlike NDVI, which can become less sensitive in lush, high-biomass environments, EVI incorporates the blue band to correct for residual atmospheric effects and is particularly responsive to variations in canopy structure and leaf area. This makes it especially effective in tropical regions like the Amazon, where dense foliage and complex vegetation layers can reduce the reliability of simpler greenness indices.
In this study, EVI was used as a tie-breaker when NDVI, SWIR, and RENDVI produced conflicting interpretations. For example, in sites such as Plantation and Regrowth, NDVI suggested greening while SWIR indicated drying, raising uncertainty about whether changes reflected healthy regrowth or stressed vegetation. Here, EVI helped clarify the interpretation: moderate to high EVI values supported the idea of vegetation recovery, even in cases where RENDVI remained low. Conversely, in locations like Passionfruit Vines and Resort Fire, EVI values near or below zero aligned more closely with RENDVI, suggesting minimal regrowth or lingering canopy stress.
However, some limitations were noted. Because PlanetScope imagery was used in unscaled digital number (DN) format rather than top-of-atmosphere reflectance, the accuracy of EVI was reduced in some areas, leading to flatter or inconclusive values. Additionally, visual banding in the mosaicked scenes may have introduced inconsistencies that impacted pixel-level results. Despite these issues, EVI generally supported field-based interpretations and proved useful in resolving ambiguity, particularly in transitional landscapes undergoing partial regrowth or degradation.
Overall, EVI enhanced the interpretive power of this multi-index approach and demonstrated its value in distinguishing between true recovery and degraded regrowth. With proper reflectance scaling and preprocessing, future applications of EVI in Amazonian land monitoring could yield even more reliable insights into forest health and change dynamics.
Field-Based Photo Validation
Field-based photo validation provided essential ground context for interpreting satellite-derived NDVI and SWIR changes, enhancing the reliability of land cover assessments. In most locations, spectral signatures corresponded well with visible land use types, particularly where vegetation recovery, degradation, or disturbance was distinct. For instance, sites like the water buffalo farm and new plantation exhibited strong agreement across NDVI, SWIR, and RENDVI, consistent with lush, moist conditions or early regrowth visible in photographs. Similarly, areas undergoing degradation, such as recently cleared or recovering land, showed expected patterns of decreasing NDVI and increasing SWIR, supported by sparse canopy or visible soil in field imagery.
However, not all locations showed full alignment. The passionfruit farm, while highly vegetated, revealed structured rows with visible ground gaps that may have led to a lower canopy signal than expected, resulting in partial agreement. At the manioc clearing and burned resort area, discrepancies in RENDVI values pointed to possible image noise or spatial averaging effects in PlanetScope data. In these cases, high local variability or unresolved fine-scale features may have affected classification consistency. Overall, these visual validations helped explain outliers and supported the interpretation that spectral change analysis is effective for general land use transitions but may benefit from additional ground-truthing when classifying mixed-use or transitional areas.
Limitations
Several constraints affected the scope and interpretation of this study. Most notably, the Landsat imagery and PlanetScope imagery did not cover the exact same geographic extent. As a result, no direct quantitative comparison could be made between the NDVI–SWIR change analysis and RENDVI outputs. Time limitations further restricted efforts to reconcile these datasets. The PlanetScope analysis relied on a single-date image, which captures only a snapshot of vegetation conditions and does not account for seasonal or interannual variability. Although cloud cover was present in the scene, it did not obscure any of the eight selected field sites.Ground-based validation was also limited by seasonal conditions.
Fieldwork was conducted during the high-water period of the Amazon River, which restricted access to upland sites and areas that were not directly accessible by boat. As a result, ground truthing was biased toward floodplain or river-adjacent vegetation, potentially underrepresenting inland land cover variability. Additionally, the sample size of eight sites, while diverse in land use types, may not fully reflect the broader spatial heterogeneity of the study region.
Differences in spatial resolution between PlanetScope (3 m) and Landsat (30 m) imagery introduced challenges when comparing vegetation indices. Finer-scale Planet imagery may capture localized disturbances or small-scale clearings that are not visible in coarser-resolution datasets. Lastly, while RENDVI proved useful in distinguishing vegetation conditions across the PlanetScope scene, its classification thresholds were based on general literature values rather than locally calibrated data. This introduces some subjectivity and may affect the interpretation of canopy health in this specific ecosystem. RENDVI should therefore be seen as a useful supplement to NDVI-based time series, but not a substitute for longer-term trend analysis.
Although PlanetScope’s 8-band imagery provides enhanced spectral detail for land cover analysis, this study faced limitations due to incomplete spatial coverage in Bands 7 (NIR) and 8 (NIR2). These near-infrared bands are essential for calculating indices like NDVI and for detecting moisture conditions and canopy structure. Their partial coverage meant that these indices could not be reliably applied across the entire study area. While RENDVI, which uses Band 5 (Red) and Band 6 (Red Edge), remained usable due to full coverage in those bands, the absence of consistent NIR data reduced the ability to conduct comprehensive vegetation and moisture assessments using PlanetScope alone. This limitation necessitated greater reliance on lower resolution Landsat data for broader trend analysis.
Both the RENDVI and EVI raster layers used in this study were generated by mosaicking multiple PlanetScope images captured on different dates. This introduced visible banding artifacts, where seams between scenes display inconsistencies in reflectance values. These variations can affect vegetation index outputs by artificially increasing or decreasing pixel values across scene boundaries, potentially skewing interpretation of vegetation health. While RENDVI is particularly sensitive to canopy density and structure, and EVI is designed to minimize atmospheric effects, neither index is immune to radiometric inconsistencies caused by mosaicked scenes. Due to time constraints, cross-scene normalization or atmospheric correction was not applied, limiting the reliability of comparisons across image strips. These issues should be considered when interpreting spatial patterns, particularly near visible seams, and highlight the need for improved preprocessing in future analyses.