The influence of space-based technology has grown significantly in informed decision-making, especially in fields like forest resource management. Remote sensing, in particular, has provided a well-organized and layered framework for evaluating and tracking forest conditions, taking advantage of sensors with varying spatial and spectral characteristics [1].

Forest ecosystems provide humans with critical resources, including carbon sequestration, biodiversity conservation, hydrological regulation and climate stabilization. Monitoring changes in forest cover is therefore essential for sustainable forest management, particularly in ecologically sensitive mountainous regions such as the Himalayas [2]-[5].

Mapping forest extent and tracking its changes are among the most widely applied uses of satellite imagery [6][7]. Forest monitoring with the help of satellite technology enables efficient assessment of spatial and temporal changes in ecosystem structure and land cover dynamics [6][8]-[10].

Advances in remote sensing and GIS technologies have significantly improved forest monitoring by enabling rapid and economical assessment of land cover dynamics across large spatial extents [11]. The availability of high-resolution multispectral satellite datasets such as Sentinel-2 has further enhanced vegetation classification and forest change detection studies.

However, achieving high levels of accuracy has often been limited by a variety of factors like spatial, spectral, and temporal limitations of widely used public datasets like Landsat and MODIS. The launch of the Copernicus Programme by the European Space Agency and the European Union marked an important advancement in this regard. In particular, Sentinel-2, operational since 2015, has improved forest mapping by providing high-resolution multispectral imagery with high frequency, enabling more detailed and reliable classification across larger areas [12]-[18].

Digital image classification techniques are widely used in remote sensing for land cover mapping and change detection. Supervised and unsupervised classification approaches enable the identification of spectrally distinct land cover classes from multispectral imagery. Some of the most popular machine learning algorithms for this include minimum distance to mean, parallelepiped, and Gaussian maximum likelihood classifier [19]-[21]. ISODATA clustering was used to group spectrally similar pixels with Maximum Likelihood Classification used to refine the final land cover categories.

Change detection is essentially the comparative analysis of two or more remotely sensed images acquired over the same geographic area across different time periods [10]. The classification of spatial data has seen a marked rise in adoption in recent years due to its capacity to uncover meaningful patterns and generate actionable insight from complex datasets. This trajectory has been shaped by extensive empirical testing across a wide range of classifiers, leading to measurable gains in classification accuracy and overall performance. Land cover mapping through satellite image classification constitutes one of the primary remote sensing applications examined in this study. Although numerous studies have applied Sentinel-2 imagery for land cover mapping yet they do not account for forest canopy transition dynamics in rapidly urbanizing Himalayan municipal forests or an integrated hybrid classification workflow. Most previous studies emphasize classification accuracy or overall land cover statistics while limited attention has been given to quantifying inter-class forest transitions using efficient statistical approaches suitable for operational efficiency, as individual classes carry very important context. Table 1, as indicated below, presents an overview of related investigations that utilized a variety of remote sensing techniques and data sources, highlighting the conclusions that proved instrumental in guiding forest cover change classification and estimation.

Table 1. Compilation of tools and datasets for analyzing changes across regions.

Monitoring forest cover dynamics in Himalayan ecosystems is particularly important due to increasing urban expansion, tourism activity, and climate changes. Shimla represents a rapidly urbanizing mountain landscape where forest fragmentation and canopy modification can significantly influence slope stability, biodiversity, water cycles, and local climate regulation. The integration of open source Sentinel-2 imagery with statistically robust classification methods provides an efficient and scalable framework for long-term forest monitoring in mountainous environments where extensive field surveys are often difficult to conduct.

This study contributes to the existing literature by introducing an integrated remote sensing workflow for forest monitoring that combines NDVI pre-processing and ISODATA spectral clustering with Maximum Likelihood Classification using high-resolution Sentinel-2 imagery. Unlike many previous studies that limit scope to overall land cover changes, this work emphasizes detailed canopy density transition analysis across five forest classes through a transition matrix framework. The methodology achieves high classification accuracy while relying exclusively on freely available satellite imagery and widely accessible GIS tools, which is efficient and reproducible solution for forest monitoring in complex mountainous terrain like the Himalayas. Furthermore, the study establishes an updated baseline dataset for Shimla Municipal Forests that can support future ecological assessments, sustainable forest management and climate adaptation planning across the region.

Multispectral Sentinel-2 imagery provided by the European Space Agency was utilized for both classification and change detection, offering a spatial resolution of 10 meters and a wide swath width of approximately 290 km. The datasets for the study area were obtained from the Copernicus open-access platform for the dates 20 November 2015 and 6 November 2019. Sentinel-2 multispectral imagery with 10 m spatial resolution and five-day revisit frequency was used for forest classification and change detection. Sentinel 2 with Level 1C top of the atmosphere reflectance products were used. Bands B2 (Blue), B3 (Green), B4 (Red), and B8 (Near-Infrared) were selected at 10 m resolution. Cloud and shadow-contaminated images were excluded during image selection. Atmospheric correction, geometric pre-processing and terrain correction were performed in ERDAS Imagine

Figure 1. Geographic location of Shimla Municipal Forests within Himachal Pradesh, India.

using the Shuttle Radar Topography Mission (SRTM) 30 m Digital Elevation Model (DEM). Images acquired during November 2015 and November 2019 were selected to minimize seasonal variability and ensure consistent comparison.

Area statistics in the results section were normalized to the total surveyed extent of the Shimla Municipal Forest boundary, 2780.4 ha rather than computed directly from pixel count at nominal 10 m resolution. This normalization accounts for boundary clipping effects at the edge of the irregular study area polygon.

Figure 2 shows a detailed framework for forest cover mapping and steps involved to generate forest maps as follows.

The methodological novelty of this study lies in integrating unsupervised spectral clustering with supervised statistical classification within a unified workflow that improves efficiency of assessment of forest canopy density classes while maintaining computational efficiency and reproducibility for regional forest monitoring applications. The methodology is based on digital Image processing

Figure 2. Methodological framework adopted for forest cover mapping and change detection using Sentinel-2 imagery and maximum likelihood classification.

and classification using ERDAS Imagine software [32] that involves the use of NDVI transformation, unsupervised hybrid classification with ISODATA algorithm and Maximum Likelihood Estimation (MLE) algorithm. The sequential steps included: 1) Data preparation (pre-processing) by downloading images, 2) ortho-rectification, 3) image interpretation, 4) categorization into layers, 5) interpretation of images to identify change areas and 6) map generation.

Both 2015 and 2019 images were ortho-rectified to a common reference system using ground control points and a digital elevation model. Registration achieved a root mean square error of 0.31 pixels within the 0.5 pixel threshold recommended for reliable change detection. This reduced false transitions between adjacent canopy classes on the steep slopes used in Section 3.

The Normalized Difference Vegetation Index (NDVI) is one of the most commonly used indicators for assessing vegetation condition, as it captures both the presence and health of vegetation by utilizing the Near-Infrared (NIR) and red portions of the electromagnetic spectrum. It is computed using the following expression:

NDVI values range from −1 to +1, where higher values indicate dense vegetation and lower values represent sparse or non-vegetated surfaces. The index was used to separate vegetated and non-vegetated areas prior to classification [22][33][34].

Unsupervised classification, which can be best described as a clustering approach because of it being largely an automated technique in which pixels with similar spectral characteristics are grouped together without prior training data. ISODATA (Iterative Self-Organizing Data Analysis Technique) is one of the most commonly used algorithms for vegetation analysis in satellite imagery. This method organizes pixels into clusters based on spectral similarity and is typically applied to separate forest classes from other land cover types. Pixels are repeatedly reassigned through an iterative process, class statistics are updated, and class boundaries are refined until stable patterns in the spectral data begin to emerge.

In this study, the vegetated portion of the imagery is classified using the ISODATA algorithm to generate categories such as Very Dense Forest (VDF), Moderately Dense Forest (MDF), Open Forest (OF), scrubland, and non-forest areas, including agriculture, land built by humans, and sparse vegetation. The algorithm operates through repeated cycles of clustering and adjustment, gradually improving the separation between land cover types based on their spectral behavior. ISODATA clustering was iteratively applied to group spectrally similar pixels into preliminary land cover classes using standard convergence and splitting criteria. Rather than using ISODATA as the final classification output the spectral clusters were subsequently refined through Maximum Likelihood Classification, which reduces spectral ambiguity between adjacent vegetation classes.

Maximum Likelihood Estimation (MLE) is a supervised parametric classifier that assigns pixels to the class with the highest statistical probability based on class mean vectors and covariance matrices [35]. MLE is a supervised, parametric classification method in which the analyst first defines representative training samples, often referred to as training areas or regions of interest. These samples are used to derive statistical descriptions of each class, which are then fed into the classification algorithm. The model computes probability density functions for each class and uses them to determine the likelihood of a pixel belonging to a given category.

Satellite image classification in this study was carried out using digital image processing techniques based on the Maximum Likelihood Estimation (MLE) approach. A pixel-by-pixel comparison framework was adopted to generate change information, enabling a more precise interpretation of land cover transitions through a clear “from-to” analysis of pixel categories. To quantify changes over time, classified outputs from two different dates were compared using a cross-tabulation method. This allowed the extent and nature of land cover change to be determined using Sentinel-2 datasets acquired on 20 November 2015 and 6 November 2019. For interpretation purposes, forest cover was mapped into canopy density-based classes. In total, five forest cover categories were defined, as outlined in Table 2.

Table 2. Forest cover classified in terms of canopy density classes.

The canopy density thresholds in Table 2 follow Kumar et al. (2007) [22]. NDVI values were binned into density ranges matching these thresholds. ISODATA clusters falling within each range were labeled Non Forest, Open Forests, Moderately Dense Forest, Scrub or Very Dense Forests to generate training areas for Maximum Likelihood Classification. The set of spectral signatures derived from unsupervised classification is stored for each image scene and subsequently used in the application of Maximum Likelihood Estimation (MLE). This approach is particularly suitable because it performs well with input data that follows a normal distribution and incorporates both variance and covariance information, making it statistically robust.

Bayesian probability theory used by the classifier evaluates each pixel by comparing it to class-level statistical parameters. Mean vector and covariance matrix derived from training data form the basis for computing class membership probabilities. A pixel is ultimately assigned to the class for which it has the highest likelihood of belonging.

In Equation (2), D_c represents the likelihood distance for class c, where X is the pixel feature vector, M_c is the class mean vector, and Cov_c is the covariance matrix derived from training samples.

The classification decision is made by assigning each pixel to the class that yields the smallest value of D, which indicates the highest likelihood of membership. When prior probabilities differ among classes, then a weighting factor can be introduced to reflect this imbalance. This modified form of the Maximum Likelihood approach is commonly referred to as the Bayesian decision rule.

A hybrid classification approach was adopted. This hybrid strategy combines the strengths of unsupervised and supervised classification by leveraging spectral clusters while using class probability distributions during final classification. Initially, ISODATA unsupervised classification was applied to identify spectral clusters and generate preliminary land cover groups. These spectral signatures were subsequently refined and used as training samples within a supervised Maximum Likelihood Classification framework to produce the final classified maps. A total of 250 training samples were delineated per image, which is 50 per class in 2015 and 2019. Samples were selected from ISODATA clusters with the highest spectral homogeneity and checked against high-resolution reference imagery. Training areas were distributed across representative slope, aspect and elevation zones. ISODATA clusters were labeled based on spectral characteristics, NDVI results, reflectance signatures for Table 2 and visual interpretation before serving as Maximum Likelihood Classification training input.

The outcomes of the classification analysis are presented in Figure 5 and Figure 6. Table 3 and Table 4 indicate that the study area exhibits a mixed distribution of forest cover classes for 2015. Very Dense Forest (VDF) accounts for 19.58% of the total area (544 ha), while Moderately Dense Forest (MDF) occupies the largest forested portion at 27.37% (760 ha), Open Forest covers 16.13% (448 ha), and Scrub Forest represents a small fraction of 1.15% (32 ha) and non-Forest land constitutes 35.53% (988 ha) of the total geographical extent. Overall, the Shimla Municipal Forest spans approximately 2780 hectares, with a substantial share falling outside forested categories. The 2780 ha figure refers to the total classified image extent used for change detection.

Figure 3. False Color Composite (FCC) Sentinel-2 imagery of the Shimla study area for 2015.

Figure 4. False Color Composite (FCC) Sentinel-2 imagery of the Shimla study area for 2019.

Figure 5. Classified forest cover map of Shimla Municipal Forests for 2015 with corresponding area statistics.

Figure 6. Classified forest cover map of Shimla Municipal Forests for 2019 derived from Sentinel-2 imagery.

Table 3. Forest cover Shimla Municipal Forest corporation in 2015 (in ha).

Statistical analysis calculated from Figure 5.

Table 4. Forest cover municipal forest/municipal corporation in 2015 (in ha).

Shimla Municipal Corporation Area = 3500 ha, Shimla Municipal Forest Area = 2780 ha.

Common conversion factor is applied for extrapolating Shimla Municipal Forest area (2780 ha) to Shimla Municipal Corporation area (3500 ha).

Common Factor = 3500/2780 = 1.25899.

Municipal Corporation area estimates were derived by multiplying Municipal Forest area values by the common factor. Class areas are scaled by a factor of 0.0122 ha per pixel, which is derived from the total classified pixel count of 227,432 in Table 3 and Table 5 against the known municipal forest boundary area rather than the nominal 100 m² per pixel Sentinel 2 resolution.

Table 5. Forest cover Shimla Municipal Forest/corporation 2019 (in ha).

Statistical analysis calculated from Figure 6.

Comparison between 2015 and 2019 indicates an increase in VDF and MDF categories, whereas Open Forest area declined during the study period (Tables 5-8). Figure 6 shows the outcome of the analysis for the year 2019.

Table 6. Forest cover municipal forest/municipal corporation 2019 (in ha).

Table 7. Forest cover of Shimla Municipal Forest in 2019 (in ha).

Table 8. Area and amount of change in forest cover in municipal forest between 2015-2019 (in ha).

Between 2015 and 2019, VDF and MDF increased while OF decreased, indicating canopy densification and possible vegetation recovery. The 2019 forest cover distribution is given in Table 7.

The change detection between 2015 & 2019 shows that there is an increase of 62 ha in VDF, 76 ha in MDF but decrease of 113 in OF, thus an increase of 25 ha among three categories, Scrub also increased by 25 ha but non-forest shows decrease of 48 ha in Shimla Municipal Forest (Table 8).

Similarly, change detection observed in Shimla Municipal corporation shows an increase of 78 ha VDF, 96 ha MDF and decrease of 142 ha of OF, though Scrub increases by 29 ha but non-forest decreases by 69 ha as discussed in Table 9.

Table 9. Area and amount of change in forest cover in municipal corporation between 2015-2019 (in ha).

Table 10 presents the forest cover transition matrix between 2015 and 2019.

Table 10.Forest cover change matrix for Shimla Municipal Forests between 2015 & 2019 (in ha).

Net change refers to the overall difference in forest cover between two time periods, obtained by combining both gains and losses observed across the study area during the two assessments.

Classification accuracy was evaluated using stratified random sampling and confusion matrix analysis with results presented in Table 11. A total of 250 reference validation points were generated across the five land cover categories using high-resolution images, Sentinel-2 true color composites, and existing forest cover maps. Accuracy assessment followed standard remote sensing validation procedures overall accuracy, and the Kappa coefficient.

Table 11. Confusion matrix and classification accuracy assessment for forest cover classification.

The overall classification accuracy achieved was 89.2%, while the Kappa coefficient was calculated as 0.86, indicating substantial agreement between classified outputs and reference data. Producer’s accuracy values ranged from 80% to 90%, whereas user’s accuracy values ranged from 82% to 91% across the five land cover categories. The results demonstrate that the hybrid ISODATA and MLE framework produced reliable classification performance for mountainous Himalayan terrain. This accuracy value falls within the acceptable range reported for Sentinel-2-based forest classification studies conducted in heterogeneous mountainous environments, where overall accuracies commonly range between 80% and 90%.