Ethnic-Led Forest Recovery and Conservation in Colombia: A 50-Year Evaluation Using Semi-Automatic Classification in the Tucurinca and Aracataca River Basins

Abstract

Deforestation in Colombia, driven by armed conflict and illicit crops, triggered an environmental crisis, particularly in the Caribbean region, where forest loss in areas such as the Sierra Nevada de Santa Marta degraded ecosystems, reduced carbon sequestration, and increased soil erosion, threatening biodiversity and local food security. In response, the Arhuaco Indigenous community implemented an ethnic territorial management system to restore degraded lands and safeguard their ancestral territory. This study evaluates the effectiveness of their efforts, supporting their call for territorial expansion by analyzing forest cover changes (1973–2023) in the Tucurinca and Aracataca river basins. Using Landsat imagery, remote sensing, and a maximum likelihood algorithm, we generated thematic maps and statistical vegetation change data, validated by a 91.4% accuracy rate (kappa coefficient and confusion matrices). Results demonstrate significant forest recovery, highlighting collective reforestation and Indigenous sustainable management as pivotal strategies for reversing deforestation in post-conflict scenarios.

1. Introduction

The armed conflict had a crucial influence on deforestation in Colombia [1]; its origin is given by the persistent inequalities [2], land tenure disputes [3], the boom in illicit drugs [4], and, in general, the struggle for resource control [5]. These factors hinder the construction of positive peace and an environment conducive to development, negatively impacting soil quality, water resources, and the provision of ecosystem services, particularly for rural populations [6]. Consequently, in recent years, there has been growing awareness of the various ways armed conflicts can affect forest conservation [7]. It is understood that the relationship between conflict and deforestation can manifest in multiple ways: conflict may reduce deforestation when armed groups use forest cover for concealment [8], or it may exacerbate environmental degradation when these groups finance their operations through the exploitation of natural resources, promoting unsustainable land uses [9,10,11].

These dynamics generated pressure on ecosystems, with the Caribbean region presenting the greatest loss of forest cover in proportion to other regions of Colombia [10], as it has less tropical forest cover (only 3%), yet it concentrated 14% of the deforested area in 2014 [12], one year before the ceasefire that would open the door to the peace agreement. This agreement sought the eradication of illicit crops and the development of policies directly focused on forest conservation and natural resource management within the framework of the post-conflict period [11], becoming a historic opportunity to lessen the impacts caused by violence in the region, which accounted for 23% of the nation’s massacres and left the highest proportion of abandoned land [13].

One of the key aspects of the region in marking this wave of violence is its vast diversity of ecosystems, most of which are found in tropical areas and can be categorized into three major groups: terrestrial, marine aquatic, and freshwater aquatic ecosystems [14]. This led to the presence of various organized armed groups who compete for control over strategic geographic points, road infrastructure, and natural corridors used for drug and weapon trafficking [13,14].

Among the most relevant conflicts, it is recounted that by the early 1970s, marijuana was already being cultivated in the lower parts of the Sierra Nevada de Santa Marta for commercial purposes due to the ease of trafficking it to the United States [15]. By the late 1980s, the crackdown on marijuana crops and growers led to the transition to cocaine cultivation and drug trafficking in the area, generating unprecedented levels of violence [16], and around 1985, the different structures and fronts of the Fuerzas Armadas Revolucionarias de Colombia (FARC), Ejército de Liberación Nacional (ELN), and Ejército Popular de Liberación (EPL) formed in the region, aiming to control drug production and trafficking through extortion and security fees. It was not until 2004, with the assassination of Mamo Mariano Suárez by the FARC, that a project was launched to mitigate these impacts and restore areas of the Sierra Nevada de Santa Marta, recognizing its importance for the preservation of many species, endemic taxa, migratory species, and the distribution of water that originates there [17].

Over the past five decades, Colombia progressively developed a legal and institutional framework aimed at environmental protection and natural resource management. Beginning with Law 23 of 1973 and the National Code of Renewable Natural Resources (Law 2811 of 1974), the country established increasingly robust mechanisms for environmental governance, culminating in the creation of the Ministry of the Environment and the National Environmental System (SINA) through Law 99 of 1993. More recent policies, such as Law 2173 of 2021, emphasize ecological restoration, while Decree Law 4633 of 2011 recognizes the territorial rights and environmental stewardship of Indigenous communities. These legal instruments provide critical support for conservation efforts and frame the socio-political context of the forest transformations examined in this study.

In this context, in 2006, Gunmaku was born in the basin of the Tucurinca and Aracataca Rivers, a Talanquera village that shelters the Arhuaco community, created to protect their way of life, water sources, and biodiversity within the framework of the post-conflict period, allowing for the reinforcement of natural ecosystem protection and the promotion of sustainable development [18] through traditional practices of sustainable resource use and a deep respect for nature based on their culture [19], controlling deforestation in pursuit of simultaneously bringing empowerment, economic income, land security, peace, and conservation [20].

Given the complex interaction between armed conflict, deforestation, and reforestation efforts led by Indigenous communities in the Sierra Nevada de Santa Marta, this research aims to evaluate the spatial and temporal dynamics of vegetation cover over the past five decades, with particular attention to Indigenous-led conservation initiatives. The study seeks to provide scientific evidence to guide conservation policies and sustainable development strategies in the post-conflict context.

The main objective of the study is to analyze changes in vegetation cover between 1973 and 2023, a year that saw significant deforestation reduction in the region [21] across the Tucurinca and Aracataca river basins. Using satellite imagery and semi-automatic classification techniques (a method that provides advanced analytical capabilities and has been applied in environmental monitoring, risk assessment [22], and natural landscape change detection [23]), this research contributes to an increasingly necessary interdisciplinary approach for addressing complex issues such as environmental degradation in conflict-affected areas. Furthermore, it generates scientifically reliable data to evaluate environmental development programs through historical land use analysis and its transformations over time [24]. This methodology serves as an indispensable tool for forest loss monitoring, as it enables near real-time identification and quantification of changes in forest cover [25].

2. Materials and Methods

2.1. Study Area

The municipality of Aracataca is located northeast of the Magdalena department, Colombia, on the western slope of the Sierra Nevada de Santa Marta, approximately 88 km from the Touristic, Cultural, and Historic District of Santa Marta. It has a warm climate with an average temperature of 30 °C and an approximate elevation of 40 m above sea level. Within its political-administrative division, it is defined by nine townships, 29 rural areas, and four Indigenous-populated centers in its rural area named Yechuikin, Serankua, Dwanawimaku, and Gunmaku. These settlements were established as territorial refuges and demarcation zones during periods of conflict, where, in addition to waiting for calm to return, they also served as barriers to demarcate Indigenous territory.

As shown in Figure 1, the study area is located within the proposed expansion of the Arhuaco Reserve, a territory comprised of a network of interconnected sacred sites, recognized by the Arhuaco community as ecological–spiritual nodes [25]. Its central element is the Sierra Nevada de Santa Marta, the world’s highest coastal mountain range (5775 m) and a biosphere reserve [17]. The study area is framed between 10°37′–10°40′ N and 73°55′–74°51′ W, bounded to the north by the Tucurinca River, to the east by the current limits of the reserve, to the south by the Aracataca River, and to the west by a monument marking the boundary with a military training center, encompassing an approximate area of 13,438 ha.

2.2. Data Acquisition and Research Framework

To analyze changes in vegetation cover in the Tucurinca and Aracataca river basins between 1973 and 2023, a semi-automatic classification approach based on remote sensing and geographic information systems (GIS) was employed. These are proven tools for assessing changes in land use and land cover, aiding planners in promoting sustainability [26]. This method allows for the classification of land cover by training an algorithm with samples of spectral signatures from various materials [22].

In general, the technique used is based on the classification of Landsat program images from four different time periods at the pixel level. Spectral signatures of specific classes are calculated and compared with nearby random pixels [27], resulting in an efficient classification. This is followed by a rigorous accuracy assessment and validation process, as illustrated in Figure 2. This methodological approach is grounded in previous research, such as that conducted by [28], where images from the same system were used to analyze changes in tree cover. This contributes to a better understanding of forest dynamics in areas of social ownership, with the aim of guiding local public policies for forest management.

Thus, for the proper acquisition of the images, the polygon corresponding to the study area was initially delineated in .shp format. This polygon was then uploaded to the EarthExplorer portal (https://earthexplorer.usgs.gov/, accessed on 13 August 2023) of the United States Geological Survey (USGS) [29,30], which is responsible for the primary operation of the Landsat program. The USGS hosts the largest collection of images available under rigorous calibration standards, facilitating detailed and precise analyses.

EarthExplorer, in turn, provides access to an extensive collection of images from satellite programs and features an interactive interface with advanced tools for image search using specific filters, such as dates, geospatial location, sensor parameters, and cloud cover percentage. From this portal, four images were exported, the characteristics of which are presented in

Table 1. Main characteristics of the images used.

Date of Image Capture	Satellite System	Solar Elevation	Solar Azimuth	Spectral Bands Used
2 January 1973	Landsat 1	43.38636	135.25246	Band 4 visible green (0.5 to 0.6 µm)
				Band 5 visible red (0.6 to 0.7 µm)
				Band 6 near-infrared (0.7 to 0.8 µm)
				Band 7 near-infrared (0.8 to 1.1 µm)
27 June 2000	Landsat 7	48.77012	143.25849	Band 1 blue (0.45–0.52 µm)
				Band 2 green (0.52–0.60 µm)
				Band 3 red (0.63–0.69 µm)
				Band 4 near-infrared (0.77–0.90 µm)
				Band 5 shortwave infrared (1.55–1.75 µm)
				Band 7 mid-infrared (2.08–2.35 µm)
16 July 2013	Landsat 8	62.25619	63.13586	Band 2 blue (0.450–0.51 µm)
				Band 3 green (0.53–0.59 µm)
				Band 4 red (0.64–0.67 µm)
18 June 2023	Landsat 8	61.35566	107.54024	Band 5 near-infrared (0.85–0.88 µm)
				Band 6 SWIR 1 (1.57–1.65 µm)
				Band 7 SWIR 2 (2.11–2.29 µm)

. According to the parameters specified on the portal, these images have a spatial resolution of 30 m, do not exceed 10% cloud cover, are georeferenced with the WGS84 global geodetic system, and correspond to a similar seasonal timeframe [24,31].

2.3. Preprocessing

As a preliminary phase to the development of the model, the quality of the satellite images was enhanced through atmospheric and radiometric corrections. This was carried out to minimize atmospheric effects and normalize surface reflectance, ensuring that variations in image brightness were solely due to the reflective properties of the surface and not to atmospheric conditions or variable instrumental configurations.

This process is crucial for enabling the accurate calculation of vegetation indices, which directly depend on reliable reflectance values across different spectral bands. Furthermore, it not only increases the precision of the analyses, but also enhances the model’s ability to detect subtle changes in land cover and identify vegetation patterns over time.

Subsequently, the images were cropped based on the region of interest (ROI), isolating only the study area. This step removed irrelevant information, optimized computational performance, and improved the relevance of the results by excluding unrelated data.

Radiometric and Atmospheric Correction of Bands

To optimize the model training, a radiometric correction of the digital numbers (DN) recorded by the satellites was initially performed, converting them into physical units, specifically into radiance values (W/m²/sr/µm) and top-of-atmosphere (TOA) reflectance. This process allows for adjusting the reflectance or brightness of certain satellite images based on the scene’s illumination [32]. This adjustment considers the amount of solar energy incident on the Earth’s surface and the lighting geometry, thus eliminating the variations caused by the position of the sun through the following equation:

$${\rho }_{TOA}={\displaystyle \frac{\pi \cdot L_{\lambda }\cdot d^2}{E_{sun}\cdot \cos \left({\theta }_s\right)}}$$

(1)

where

• ${\rho }_{TOA}$ = the TOA reflectance.
• $L_{\lambda }$ = the radiance measured by the sensor at a specific wavelength.
• $d$ = the Earth–Sun distance in astronomical units (AU), which varies slightly throughout the year due to Earth’s elliptical orbit.
• $E_{sun}$ = the extraterrestrial solar irradiance at the corresponding wavelength, representing the amount of solar energy received outside Earth’s atmosphere.
• ${\theta }_s$ = the solar zenith angle.

Subsequently, a sensor sensitivity adjustment was performed to optimize the radiometric quality of the images, enhancing contrast and maximizing detail under various lighting conditions. Satellites such as Landsat use adjustable gain settings to optimize the dynamic range of images, depending on illumination conditions and scene characteristics. By converting radiance to TOA reflectance, this influence is normalized, ensuring that differences in image brightness are solely due to the reflective properties of the Earth’s surface and not to variable instrumental configurations [33].

This approach allows for data standardization, making them comparable across different acquisition dates and conditions, ensuring that satellite images maintain an optimal dynamic range, maximizing detail and improving accuracy in land cover change analysis and environmental monitoring.

This method is adopted by the USGS for atmospheric correction, based on the moderate resolution atmospheric transmission (MODTRAN) radiative transfer model, which calculates atmospheric transmittance along a specific path, accounting for absorption and scattering caused by gas molecules and aerosols such as fog, snow, and rain [23,34]. To minimize these effects, the model employs nearest neighbor interpolation within an acceptable root mean square error, eliminating absorption and scattering effects caused by suspended atmospheric particles from the radiance received by the sensor. This process converts radiance into surface reflectance values, enabling the removal of dark objects through atmospheric correction via the SCP plugin in QGIS version 3.28, thereby providing more accurate and comparable data for multitemporal analysis [35].

2.4. Training

At this stage, the field-recognized classes and their specific characteristics were defined by combining spectral bands to construct multispectral images. Various band combinations were used, along with the RGB color model for natural color (red, green, and blue), including near-infrared, red, and green for general vegetation, SWIR 1, near-infrared, and red for vegetation analysis, near-infrared, SWIR 1, and blue for healthy vegetation analysis, and SWIR 1, near-infrared, and red for land use. These combinations allowed for a clearer and more precise identification of objects and features in the satellite images.

Additionally, to ensure a robust supervised classification analysis considering that the results depend on the user defining the spectral signatures of known categories [36], a preliminary field visit was conducted. This visit aimed to recognize and interpret land cover types on-site, strengthening the analysis of aerial photographs and satellite images [37]. To accurately define the respective land cover classes, each class was analyzed in the field, as shown in Figure 3, which provides a reference for each category based on points collected during field surveys in the study area.

Subsequently, the characteristics of small regions of interest were determined for each identified land cover class in the images. To achieve this, polygons were drawn over homogeneous areas of the image, overlapping pixels belonging to the same class. For validation and complementary analysis, PlanetScope satellite images, which matched the temporal and spatial coverage of the Landsat images, were used due to their higher spatial resolution.

Additionally, vegetation indices, including the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and soil-adjusted vegetation index (SAVI), were calculated to highlight differences in vegetation characteristics, such as density and health. These indices were particularly useful in areas where visual identification of vegetation types was challenging, enabling a more accurate classification of different types of vegetation cover.

Once the spectral signature characteristics were labeled, the model was trained using the incremental region algorithm. This algorithm identifies pixels, such as a seed pixel, within regions of interest, considering the spectral similarity of adjacent pixels. It processes data incrementally, storing only the k-nearest neighbors for each object, making it efficient for handling large datasets and adaptable to new data without requiring a complete recalibration [38].

For this procedure, the Semi-Automatic Classification plugin was used, a free and open-source QGIS module designed for the supervised classification of remote sensing data [27]. Specific pixels matching the previously identified land cover types were assigned, and polygons were drawn around the pixels, representing each type of cover.

To estimate spectral separability and avoid redundancies in training, it was assumed that each class spectrum is unique. The Jeffries–Matusita distance was used, as it provides a more reliable criterion because, as a function of class separability, it behaves more similarly to the probability of correct classification [23,39]. Additionally, the Bray–Curtis similarity index was employed to calculate the separability between pairs of probability distributions and efficiently evaluate the training results [40,41]. In this context, if the results from equation 2 yield a value close to 0, it indicates very similar spectral signatures, while values equal to or greater than 2 indicate significant differences.

$$S\left(x,y\right)=100-\left({\displaystyle \frac{{\sum }_{i=1}^n\left|\left(x_i-y_i\right)\right|}{\sum _{i=1}^n{x}_i+\sum _{i=1}^n{y}_i}}\right)·100$$

(2)

where

• $x$ = first spectral signature vector.
• $y$ = second spectral signature vector.
• $n$ = number of image bands.

The reviews showed values ranging between 1 and 1.98, indicating that medium and high vegetation exhibited some degree of similarity. Therefore, subsequent corrections were implemented, following the methodology of interdependent visual interpretation, which minimize cartographic errors that occur during the classification process [24,42]. Finally, to analyze the changes, the coverages generated by the supervised classification of satellite images from 1973, 2000, 2013, and 2023 were superimposed, creating a cross-tabulation matrix that allowed for the identification and analysis of changes in the different categories, revealing the patterns that guided these changes in the territory.

2.5. Thematic Classification and Evaluation

At this stage, the maximum likelihood algorithm is used to classify the pixels of the analyzed image and determine the statistical probability of belonging to each of the classes [32]. This method calculates the probability distributions for the classes, based on Bayes’ theorem, and estimates whether a pixel belongs to a specific land cover class [23]. The classification is performed using the following discriminant function, calculated for each pixel of the image:

$$g_k\left(x\right)=\ln p(C_k)-{\displaystyle \frac{1}{2}}\ln \left|{\Sigma }_k\right|-{\displaystyle \frac{1}{2}}{(x-y_k)}^t{\sum }_k^{-1}(x-y_k)$$

(3)

where

• $C_k$ = coverage class $k$,
• $x$ = spectral signature vector of an image pixel,
• $p\left(C_k\right)$ = probability that the correct class in $C_k$,
• $\left|{\Sigma }_k\right|$ = determinant of the covariance matrix of the data in the class $C_k$,
• ${\Sigma }_k^{-1}$ = inverse of the covariance matrix,
• $y_k$ = spectral signature vector of the class $k$.

In general, in this equation, the unknown measurement vector is assigned to the class with the highest probability of membership, considering the variance–covariance within the class distributions. This method is particularly effective for normally distributed data, outperforming other known parametric classifiers [43]. According to [44], probability-based classification is more accurate than distance-based classifiers because it incorporates the mean and variance of the dataset into the decision rule, providing additional information that improves classification accuracy. However, the effectiveness of this algorithm largely depends on an accurate estimation of the covariance matrix, which requires enough pixels in each class training site [23].

Subsequently, the accuracy and margin of error of the classifications are evaluated using overall accuracy and the kappa coefficient. These metrics allow for assessing the accuracy of the classifiers and determining the informational value of the resulting data, utilizing confusion matrices for each of the classified images in relation to real-world data [45]. Overall accuracy represents the percentage of correctly classified pixels in an image, while the global kappa index evaluates the agreement between classified pixels and reference sample pixels for all included categories [46]. Both indicators were calculated from confusion matrices, enabling a comparison of the performance of the five classifiers evaluated.

To construct the confusion matrix, it was determined that the minimum sample size should be 196 elements based on the binomial probability statistic [47]. However, 250 elements were used per season, with a proportional amount assigned to each land cover according to its representation in the population. The sampling was conducted randomly to avoid repetitive patterns, and the points were validated using high-resolution images from SAS Planet (https://account.planet.com, accessed on 8 September 2023).

3. Results

3.1. Vegetation Cover and Land Use Maps

The first analysis of land cover changes is shown in Figure 4, which illustrates the spatial configuration of land cover and use in the years 1973, 2000, 2013, and 2023 in the Arhuaco Reserve expansion region. For 2013, polygons representing 50 Indigenously occupied properties were included, designated as “Arhuaco Properties”. Furthermore, the results vary in each period depending on the chronological events affecting the territory, such as the presence of illicit crops, subversive groups, land restitution to farmers, and the acquisition of properties by the Indigenous community.

3.2. Validation of Vegetation Maps

In the sampling phase, 250 points were randomly listed for each multitemporal layer. The number of evaluation points was adjusted proportionally to the area of each category. Using these data, a confusion matrix was created for each time period, from which omission and commission errors were calculated to assess accuracy, user accuracy (avoiding false positives), and producer accuracy (avoiding omissions) to evaluate class-level accuracy, as well as the kappa index, which assesses the reliability of a classification model [48], by measuring the agreement between predicted and validation results while adjusting for agreement expected by chance.

The results for 1973 show that water bodies and tall vegetation exhibit high classification accuracy (producer’s accuracy > 96%), which is attributable to their distinctive spectral signatures. Medium vegetation shows high reliability (user’s accuracy: 97.14%), though with misclassifications of low and tall vegetation, which are likely due to gradual rather than abrupt transitions in density, height, or vegetation structure, creating intermediate zones where spectral boundaries between categories are indistinct. As shown in

Table 2. Validation results for the year 1973.

1973	WB (Water Body)	BS (Bare Soil)	LV (Low Vegetation)	MV (Medium Vegetation)	HV (High Vegetation)	Total	User Accuracy	Commission Error
WB	27	1	0	2	0	30	0.9	0
BS	1	26	2	1	0	30	0.8666	0.1333
LV	0	0	45	5	0	50	0.9	0.1
MV	0	0	1	68	1	70	0.9714	0.0285
HV	0	0	0	2	68	70	0.9714	0.0285
Total	28	30	48	78	69	253
Producer Accuracy	0.9642	0.8666	0.9375	0.872	0.9855	Overall accuracy = 0.9023
Omission Error	0.0357	0.1333	0.0625	0.128	0.0144	Kappa index = 0.8441

, the overall accuracy (90.23%) and kappa index (0.8441) confirm the model’s robustness. However, limitations are observed in discriminating bare soil and low vegetation, which are caused by spectral similarities between soils and discontinuous, sparse, or patchy vegetation covers. Additionally, omissions in medium vegetation are detected, which are associated with shadowing effects and phenological variability not captured by the model.

The validation for the year 2000 revealed that the primary classification confusion occurred among vegetation strata, which is attributable to spectral similarity in the NIR (near-infrared). Errors in bare soil and water bodies were associated with atmospheric artifacts and soil moisture variability, respectively. However, according to

Table 3. Validation results for the year 2000.

2000	WB (Water Body	BS (Bare Soil)	LV (Low Vegetation)	MV (Medium Vegetation)	HV (High Vegetation)	Total	User Accuracy	Commission Error
WB	30	0	0	0	0	30	1	0
BS	1	27	2	0	0	30	0.9145	0.1000
LV	0	0	43	7	0	50	0.8612	0.1400
MV	0	0	0	67	3	70	0.9571	0.0428
HV	0	0	0	1	69	70	0.9857	0.0142
Total	31	30	45	75	72	253
Producer Accuracy	0.9677	0.9	0.9555	0.893	0.9583	Overall accuracy = 0.9077
Omission Error	0.0322	0.1	0.0444	0.107	0.0416	Kappa index = 0.8634

, the model achieves good overall accuracy (90.77%) and a kappa index of 0.863, providing a reliable quantitative basis. Furthermore, it is evident that the main limitations stem from known factors rather than methodological flaws.

The results for 2013, presented in

Table 4. Validation results for the year 2013.

2013	WB (Water Body)	BS (Bare Soil)	LV (Low Vegetation)	MV (Medium Vegetation)	HV (High Vegetation)	Total	User Accuracy	Commission Error
WB	30	0	0	0	0	30	1	0
BS	0	29	1	0	0	30	0.9666	0.0333
LV	0	0	45	5	0	50	0.9	0,1
MV	0	0	0	70	0	70	1	0
HV	0	0	0	1	69	70	0.9857	0.01428
Total	30	30	46	76	69	251
Producer Accuracy	1	0.9666	0.9782	0.921	1	Overall accuracy = 0.9124
Omission Error	0	0	0	0.079	0	Kappa index = 0.8914

, again demonstrate good performance in classifying water bodies and tall vegetation, with producer’s and user’s accuracies of 100% and 98.57%, respectively. Medium vegetation shows improved accuracy, though with 7.9% omissions, which is likely due to variability in leaf density. Bare soil and low vegetation maintain good results (accuracies >96% and >90%), but residual errors persist, as one case of bare soil was confused with low vegetation and five cases of low vegetation were classified as medium vegetation due to gradual transitions in vegetation cover and spectrum mixing in areas of discontinuous vegetation.

Finally, in

Table 5. Validation results for the year 2023.

2023	WB (Water Body)	BS (Bare Soil)	LV (Low Vegetation)	MV (Medium Vegetation)	HV (High Vegetation)	Total	User Accuracy	Commission Error
CA	30	0	0	0	0	30	1	0
SD	0	30	0	0	0	30	1	0
VB	0	0	48	2	0	50	0.9600	0.0400
VM	0	0	0	70	0	70	1	0
VA	0	0	0	3	67	70	0.9571	0.0429
Total	30	30	48	75	67	250
Producer Accuracy	1	1	1	0.9333	1	Overall accuracy = 0.9316
Omission Error	0	0	0	0.0667	0	Kappa index = 0.9022

for 2023, the validation results show that the model achieved its highest overall accuracy (93.16%) and a kappa index of 0.9022, reflecting excellent reliability. The water body and bare soil categories were identified without errors, with 100% accuracy. For vegetation cover, low vegetation and tall vegetation showed low commission errors (4% and 4.29%, respectively). Medium vegetation had an omission error of 6.67%, indicating a slight underestimation of this class; however, this error is relatively low and within expected ranges for environmental land use classifications.

3.3. Rates of Change

The spatiotemporal dynamics of land cover and land use classes were evaluated using cross-tabulation matrices, a tool that allows for the comparison of changes in territorial distribution over time. These matrices show how different cover categories changed across various periods, identifying transitions, losses, and gains in each. This method facilitates understanding which areas remained stable and which have undergone changes, as shown in Figure 5, enabling the assessment of landscape transformation trends, such as deforestation or natural regeneration, through the quantification of these changes.

Thus, according to the evolution of land cover from 1973 to 2023, a significant transformation in the landscape is evident, where tall vegetation has shown a progressive reduction, decreasing from 7413.128 ha in 1973 to 3223.790 ha in 2023, indicating an ongoing process of forest degradation. Similarly, low vegetation also declined considerably, dropping from 5885.218 ha in 1973 to 2910.210 ha in 2023, suggesting that some of these areas may have been converted for other uses, including illicit crops. In contrast, medium vegetation experienced an increase between 1973 (3759.161 ha) and 2000 (4376.772 ha), but then began to decline towards 2013 (2974.5 ha), indicating that large areas of intermediate vegetation may have been cleared for other uses, and it decreased even further in 2023 (1556.234 ha), which could be related to changes in vegetation regeneration or alterations in ecosystem dynamics.

Regarding other land covers, bare soil followed a fluctuating trend: in 1973 it covered 1696.879 ha, decreased by 2000, then increased in 2013 due to crops such as marijuana requiring the removal of native vegetation, leading to a rise in bare soil before crops became established and finally saw a reduction to 302.460 ha by 2023, which could be linked to restoration processes in certain areas. Water bodies remained stable over time, with values ranging between 241.219 ha in 1973 and 224.590 ha in 2023, suggesting relative conservation of these aquatic ecosystems.

3.4. Forest Change Dynamics

To assess natural resource management in Indigenous territories and understand vegetation changes, a multitemporal analysis of vegetation subclasses was conducted, showing changes in forestation under the Arhuaco community’s conservation strategy. To obtain statistics on these changes, the previously generated coverages were superimposed by creating a cross-tabulation matrix. This allowed us to identify changes in the different categories analyzed and the percentage changes in the surface area for each year under study, based on the following equation:

$$r=\left({\displaystyle \frac{1}{t_2-t_1}}\right)\cdot \mathit{ln}\left({\displaystyle \frac{A_2}{A_1}}\right)\cdot 100$$

(4)

where

• ${\mathrm{A}}_1=$ coverage of the initial temporality,
• ${\mathrm{A}}_2=$ final temporality coverage,
• ${\mathrm{t}}_1=$ year of the initial period,
• ${\mathrm{t}}_2=$ year of the final period.

The results of this calculation are summarized in

Table 6. Values of classes by time period, expressed in hectares (ha).

Subclasses	T1 (1973–2000)	T2 (2000–2013)	T3 (2013–2023)	% Change T1–T2	% Change T2–T3
Unchanged bodies of water	148.154	205.244	226.590	1.385%	1.104%
Bare flood-prone soill	8011	1975	0	−4.055%	−1.975%
Vegetation cover due to fluvial dynamics	85.050	1	0	−84.942%	−1.001%
Madrevieja	18.652	21.346	0	1.144%	21.346%
Unchanged bare soils	41.415	1695.894	1615.149	40.949%	−1.050%
Transition vegetation	1255.175	1458.731	5292.141	1.162%	−3.628%
Deforestation	2688.870	620.187	6789	−4.336%	−91.346%
Change to cultivation	1417.696	4419.170	2992.878	3.117%	−1.477%
Unchanged low vegetation	1873.209	4458	9488	−420.208%	2.128%
Unchanged medium vegetation	1985.444	2977.357	1259.353	1.500%	−2.364%
Unchanged high vegetation	5104.502	3220.814	3223.790	1.585%	−1.001%

, which shows the values of the temporality classes related to losses or gains in vegetation cover per hectare across the three established periods, revealing a complex land use dynamic over the five decades, with significant changes in vegetation cover and water bodies. The no change category in water bodies showed sustained growth (38.5% from 1973 –2000 to 2000 –2013 and 10.4% from 2000 –2013 to 2013 –2023), which could indicate greater conservation of these ecosystems or even expansion due to climatic changes or water management. Conversely, the complete disappearance of floodable bare soil and vegetation cover with fluvial dynamics between 2013 and 2023 suggests a drastic alteration in fluvial systems, possibly due to human interventions such as dams or changes in rainfall patterns.

Regarding vegetation, deforestation decreased considerably between 2013 and 2023 compared to 2000 and 2013, which could be due to conservation strategies implemented by the community or the scarcity of remaining forests available for logging. However, the change to cropland category showed a notable increase in 2000–2013 (211.7% compared to 1973–2000), followed by a decline between 2013 and 2023 (−32.3%), suggesting a possible saturation of arable land or a shift toward more sustainable practices.

The vegetation categories display contrasting trends: while tall and medium vegetation decreased between 2013 and 2023 (−1% and −57.7%, respectively), low vegetation increased (112.8%), which could indicate the replacement of mature forests or crops with newer, more sustainable or transitional crops. Finally, the no change bare soil category experienced a massive rise between 2000 and 2013 (3.995% compared to 1973–2000), followed by a slight decrease in 2013–2023 (−4.8%), potentially linked to intensive agriculture that later stabilized. Meanwhile, the vegetation transition category saw an exceptional increase in 2013–2023 (262.8% compared to 2000–2013), which might reflect ecological succession or restoration processes.

The results reveal a drastic decline in flood-prone bare soil, which decreased by 75.3% (from 8011 ha to 1975 ha) between 1973 and 2000 and disappeared entirely by 2013–2023. This reduction could be attributed to the channelization of waterways, agricultural expansion into low-lying areas, and other anthropogenic interventions.

In contrast, madreviejas increased by 14.4% between 1973 and 2013 (from 18,652 ha to 21,346 ha), which is linked to natural processes such as meandering and the disconnection of secondary river channels. However, during the 2013–2023 period, they vanished completely (0 ha), which was likely due to a loss of hydrological connectivity caused by alterations to the fluvial system.

As evidenced in Figure 6, these changes reflect a significant landscape transformation, characterized by the progressive loss of high and medium vegetation cover. This phenomenon is primarily associated with deforestation and the expansion of cropland.

The analysis included polygons corresponding to 50 properties occupied by Indigenous communities (called “Arhuaco Properties” since 2000). Between 2000 and 2013, an intensification of environmental degradation was observed, with an increase in transition areas and deforested zones likely linked to agricultural expansion. However, in the period of 2013–2023, a reduction in bare soils prone to flooding and an increase in transition vegetation were recorded. This trend suggests that strategies such as the eradication of illicit crops, conservation policies, or even land abandonment (associated with conflict or migration) could have favored the natural regeneration of vegetation cover.

4. Discussion

The 50-year vegetation analysis presented in this study reveals significant land cover transformations driven by both anthropogenic and natural processes, particularly within the socio-ecological context of the Sierra Nevada de Santa Marta. Although this study emphasizes anthropogenic drivers such as conflict-related displacement, land use change, and Indigenous governance, vegetation change in the Sierra Nevada is the result of interacting factors. These include abiotic influences such as drought and fire, and biological processes such as natural succession and river meandering. Thus, forest dynamics in the region should be interpreted as the outcome of complex socio-ecological interactions, rather than isolated causes.

Between 1973 and 2000, tall vegetation suffered a net loss of 88,049 ha, reflecting a deforestation process, while low vegetation increased by 95,457 ha, which could indicate degradation of dense forests or secondary regeneration with marijuana crops [15]. Bare soil grew alarmingly at a change rate of 6.387%, possibly due to military operations, the establishment of strategic bases, and the displacement of communities that traditionally protected the ecosystems [49].

During 2000–2013, low and tall vegetation showed signs of recovery, with increases of 165 ha and 153 ha, respectively. However, this phenomenon could be linked to the expansion of illicit crops in forested areas, as by 2001, the region faced one of the most severe events in its history: intense clashes between the Autodefensas Unidas of Colombia (AUC) and the self-defense groups of El Mamey or the Frente Resistencia Tayrona forced the displacement of nearly 9000 people toward the Santa Marta-Riohacha highway [50]. Water bodies lost 97.5 ha, of which 6 ha turned into bare soil, which was likely due to water extraction or the effects of climate change. Additionally, bare soil gained 68 ha in areas previously covered by tall and medium vegetation, suggesting deforestation, droughts, or wildfires.

Between 2013 and 2023, the analysis of the broader area reveals the disappearance of 99 ha of water bodies, primarily transformed into tall vegetation. A significant conversion of 510 ha of bare soil to low vegetation was recorded, indicating ecological regeneration. However, part of the tall vegetation growth occurred over vanished water bodies, which may reflect degradation rather than environmental recovery.

During the same period, in Indigenous-occupied lands, water bodies remained stable at 3 ha, while 235 ha of bare soil transitioned to low vegetation and 13 ha to medium vegetation, suggesting natural reforestation processes. Medium vegetation increased from 778 ha to 1277 ha, mainly due to the conversion of 124 ha of low vegetation and 13 ha of bare soil, demonstrating ecosystem recovery. Finally, tall vegetation showed notable growth from 1158 ha to 1922 ha, with minimal losses to bare soil, indicating effective conservation of existing forests.

These findings are directly related to the 2021 monitoring results of territories affected by illicit crops [51], which report a significant reduction in coca leaf and marijuana production in the Sierra Nevada region. This reduction stems from voluntary eradication efforts by local Indigenous communities, manifested through the cultivation of cacao and coffee. Likewise, the reports emphasize the importance of collective efforts within the framework of the total peace initiative [52]. As demonstrated in this research, proper territorial planning leads to development focused on restoring sustainable livelihoods. Furthermore, promoting community reconstruction progressively strengthens post-conflict governance mechanisms both in relation to nature and among communities themselves [53].

The findings of this study have significant implications for evaluating vegetation cover and its long-term environmental transformations, particularly in territories affected by armed conflicts, as in Colombia’s case. Furthermore, the analysis proved efficient in terms of time and costs due to data availability and the use of open-source software, facilitating its application in different contexts without requiring major investments. Despite the vast extension of the analyzed area, the employed methodology demonstrated effectiveness in detecting vegetation cover changes with high precision [54].

The use of approaches based on the integration of geographic information systems, Landsat imagery, semi-automatic classification, and maximum likelihood algorithms [55] enabled obtaining highly precise results. Validation through the kappa coefficient reflected an overall accuracy of 91.4%, surpassing values reported in previous studies that used supervised and unsupervised classification techniques supplemented with semi-automatic classification, which achieved accuracy of 85.47% [22], 84.3% [56], or 89.76% [23]. These accuracy improvements suggest that combining multiple tools and approaches optimizes land cover change detection, reducing analysis uncertainty and enhancing result reliability.

Beyond its accuracy, the proposed methodology stands out for its versatility and applicability across diverse environmental scenarios. Its replicability makes it a valuable tool for monitoring deforestation processes, ecological regeneration, and territorial planning in other regions with similar conditions [57]. The precise identification of ecosystem degradation and recovery patterns not only contributes to informed decision-making regarding conservation and environmental restoration, but also provides key information for designing public policies focused on sustainability and resilience in territories affected by conflicts or land use changes.

While this study provides a robust long-term assessment of vegetation cover changes, some limitations should be acknowledged. The temporal resolution, divided into broad periods (1973–2000, 2000–2013, and 2013–2023), effectively captures general trends but may miss episodic events such as wildfires, droughts, or rapid deforestation. Similarly, the spatial resolution of historical imagery (particularly from 1973) limits detailed interpretation of land cover transitions, as large pixels may blend different vegetation types and reduce accuracy in ecotonal zones. Nevertheless, our consistent classification methodology and post-classification validation ensure reliable cross-temporal comparability of results.

For future research, we recommend integrating higher-resolution imagery (e.g., Sentinel-2), drone surveys, or LiDAR data (particularly for shorter timeframes and more recent years) to better capture short-term dynamics and improve spatial detail. This approach would prove especially valuable in the Sierra Nevada de Santa Marta, where understanding the interplay between climate variability, vegetation recovery, and Indigenous governance is crucial for developing effective conservation policies. Such advancements would directly contribute to climate adaptation strategies and sustainable development in this strategic ecosystem.

5. Conclusions

The Sierra Nevada de Santa Marta, one of Colombia’s most biodiverse and culturally significant ecosystems, endured decades of socio-environmental transformations. This study evaluates vegetation cover changes (1973–2023) using Landsat imagery and maximum likelihood classification, focusing on the Tucurinca and Aracataca river basins. Results reveal how armed conflict, illicit crops, and Indigenous conservation strategies shaped patterns of deforestation and forest regeneration.

The most severe deforestation occurred between 1973 and 2000 (18.38% loss, 6.387% annual rate), and was linked to marijuana cultivation and territorial violence. Subsequently, rates declined (4.24% in 2000–2013; 1.15% annually), with a net recovery of 0.05% observed in 2013 –2023—the lowest deforestation rate of the study period. This transition aligns with the implementation of the Arhuaco governance model, which dedicates 70% of their territory to conservation through traditional ecological knowledge and sustainable practices, emerging as a benchmark for other regions.

While full ecological restoration remains challenging, these findings highlight the potential of Indigenous-led initiatives to reverse deforestation. The Arhuaco success demonstrates that combining local governance, ancestral knowledge, and strategic adaptation is crucial for conservation. This study advocates for participatory research frameworks that empower Indigenous communities to strengthen both local and global sustainability policies.