Introduction

Synthetic Aperture Radar (SAR) has improved Earth observation by providing cloud-free and independent of sun illumination and high-resolution data. Full Polarimetry Synthetic Aperture Radar (Full-PolSAR) data can be used in different applications for detailed classification and environmental monitoring. The full-PolSAR data contains both the co-polarized, i.e., HH and VV, as well as the cross-polarized, i.e., HV and VH backscatter information, respectively, which creates a scattering matrix that captures the complexity of target interactions with the radar signal providing detailed descriptions of scattering mechanisms. This improves the accuracy of the classification (Chen et al., 2018; Maitra et al., 2013; Nord et al., 2009; Wang et al., 2019; Yang et al., 2021).

It is essential to decompose the scattering mechanisms from the PolSAR data to discriminate between surface, double-bounce, and volume scattering. Various model-based and eigenvalue-based scattering power decomposition techniques (An et al., 2010; Sato et al., 2012; Singh & Yamaguchi, 2018; Singh et al., 2019; Yamaguchi et al., 2005, 2006) have been proposed to decompose the total scattering power into different components of scattering power related to different features on the Earth’s surface. The scattering power derived from these decomposition techniques can be used for accurate classification and applications in deforestation detection (Avtar et al., 2012; Priyanka et al., 2023; Wiederkehr et al., 2020), biomass estimation, and change detection (Avtar et al., 2013; Hong et al., 2019; Rajat et al., 2024; Suab et al., 2024). Full-PolSAR data have complete scattering information, which leads to high accuracy in different applications, but the widespread utilization of full PolSAR data is limited due to sparse availability, technical complexities, and higher costs in data collection and processing. (Moreira et al., 2013). Such restrictions lead to limited spatial coverage and temporal resolution compared with single- or dual-polarization data (Deng et al., 2023; Schmullius & Evans, 1997), which limits the application of full-polarimetric SAR data in large-scale and time-sensitive monitoring applications (Schmullius & Evans, 1997).

In contrast, dual-polarization (dual-pol) SAR data, which captures only two polarization channels, either co-polarized or cross-polarized, is more widely accessible than full-polarimetric (full-pol) data from various SAR satellite missions. The increased availability and broader swath coverage of dual-pol data makes it advantageous for large-scale monitoring and time-series applications, especially in regions requiring frequent and extensive monitoring, such as tropical forests subject to rapid deforestation and degradation (Hansen et al., 2020). Recognizing the potential of dual-pol data (Sugimoto et al., 2022, 2023) has extended the scattering power decomposition algorithms to dual-pol data by developing a three-component scattering power decomposition, that is, volume scattering, ground scattering, and double-bounce scattering.

The main objective of this study was to create a new dual-pol four-component scattering power decomposition model (DP-4SD). This model uses dual-polarization SAR data to improve detection and classification, including classification of forest species and disturbance detection, ecology, and natural resource management applications. In the 4SD model, a new oriented dipole scattering component was developed to incorporate scattering from different orientations of the scattering objects, including but not limited to the forest canopy or oriented urban features that are not provided by a three-component scattering model in the dual-pol data. Including the oriented dipole scattering component provides a more nuanced representation of the scattering behavior in complex environments, mainly where volume scattering dominates, such as in forested areas, because the three scattering power decomposition models (3SD) overestimate the contribution of volume scattering, resulting in less accurate classification results. The 4SD model reduces the overestimation of volume scattering, thereby improving polarimetric information extraction accuracy and target classification. The oriented dipole (OD) scattering model was selected because it effectively represents the scattering behavior of various vegetation structures when analyzed using dual-polarization (dual-pol) data. The OD scattering power captures the predominant volume scattering mechanism in vegetation areas where the scatterers (such as tree branches and leaves) are randomly oriented. This is critical for distinguishing vegetation from other land cover types, as vegetation typically exhibits strong volume scattering due to the random orientation of dipoles within the canopy (Freeman & Durden, 1998).

While other models like compound dipole or mixed dipole scattering can also represent vegetation, but oriented dipole can be generated using dual-pol data, as it captures the dominant scattering mechanism without requiring the full complexity of full-pol data. In this study, oriented dipole provides a balance between accuracy and availability of the full PolSAR data in environments dominated by volume scattering, such as forests or agricultural areas. By developing this 4SD model, this study aims to bridge the gap between the limited availability of full-polarimetric SAR data and the need for advanced decomposition techniques to effectively analyze dual-pol data. The DP-4SD model can be applied to various applications, including forest monitoring, urban planning, and disaster management, where accurate and timely information is critical for decision making using SAR-based remote sensing technologies, contributing to improved environmental and resource management globally.

Test Sites and Data Used

Haldwani Forest

The Haldwani forest range is located in between 29° 00′ N and 29° 15′ N latitude and 79° 15′ E and 79° 30′ E longitude (Fig. 1a) in the Uttarakhand state in the foothills of the lower Shiwaliks. The climate of the Haldwani Forest Range is typically subtropical, characterized by hot summers, mild winters, and a significant monsoon season. The elevation ranges from approximately 500 m (1640 ft) above sea level in the lower areas to over 1000 m (3281 ft) in the higher regions near hills. It is a managed forest region with unique species distribution in well-delineated forest compartments.

Fig. 1
figure 1

Test Sites a Haldwani Forest b San Francisco, \(\langle {\left|{S}_{VV}\right|}^{2}\rangle\), \(\langle {\left|{S}_{VH}\right|}^{2}\rangle\), and \(\langle {\left|{S}_{VV}\right|}^{2}\rangle\)/\(\langle {\left|{S}_{VH}\right|}^{2}\rangle\) are the elements of the covariance matrix

Full size image

In polarimetric synthetic aperture radar (PolSAR) analysis, creating an RGB composite image using the covariance matrix \({[C}_{2}]\) elements \(\langle {\left|{S}_{VV}\right|}^{2}\rangle\), \(\langle {\left|{S}_{VH}\right|}^{2}\rangle\) and the ratio \(\langle {\left|{S}_{VV}\right|}^{2}\rangle\)/\(\langle {\left|{S}_{VH}\right|}^{2}\rangle\)​ is an effective way to visualize different scattering mechanisms and surface properties (\(\langle .\rangle \text{represents spatial ensemble avarageing})\). Each of these elements captures specific backscattering characteristics in different polarization channels, providing valuable insights into land cover classification and environmental monitoring.

San Francisco

San Francisco, located on the western coast of the United States, is an urbanized region characterized by diverse land-use and land-cover (LULC) types, including urban areas, vegetation, water bodies, and infrastructure (Fig. 1b). As a study area, San Francisco offers a complex environment for classification analysis because of its varied topography and heterogeneous landscapes, including natural features such as hills, forests, and coastal regions, as well as dense urban structures. It has a Mediterranean-type climate, characterized by mild wet winters and warm dry summers, which supports high biological diversity.

Data Used

In this study, we utilized datasets from the Advanced Land Observation Satellite (ALOS), equipped with the Phased Array type L-band Synthetic Aperture Radar (PALSAR), and the Advanced Land Observing Satellite-2 (ALOS-2), which features the Phased Array type L-band Synthetic Aperture Radar-2 (PALSAR-2). PALSAR’s ability to operate in multiple polarization modes enables detailed surface analysis, making it effective for applications such as deforestation monitoring and target classification (Hayashi et al., 2019; Tadono et al., 2019; Zhang et al., 2023). Datasets were acquired for both test sites as mentioned in Table 1.

Table 1 Details of the PolSAR datasets
Full size table

The ALOS-2/PALSAR-2 datasets were processed to generate \({[C}_{2}]\) using a multi-look window factor of 2 in the range direction and 5 in the azimuth direction. For ALOS/PALSAR, \({[C}_{2}]\) was generated with a multi-look window factor of 1 in the range direction and 5 in the azimuth direction. To enhance the accuracy of real-world representation, geometric distortions were addressed through terrain correction applied to the \(\left[{C}_{2}\right]\). This process utilized a Shuttle Topography Radar Mission Digital Elevation Model (SRTM-DEM) with a 30 m posting resolution to compensate for any topographic-induced distortion.

Concept of Dual-Pol Model-Based Decomposition

2 × 2 Covariance Matrix

There are only two complex elements (\({S}_{VV}\) and \({S}_{VH}\) or \({S}_{HH}\) and \({S}_{HV}\)) in dual-polarization SAR measurements, the coherency matrix and Pauli scattering vector cannot be created using dual-polarization data. Based on the relationship between dual-polarization and full PolSAR data, we can summarize the polarimetric information of the 2 × 2 covariance matrix \(\langle \left[{C}_{2}\right]\rangle\) ( \(\langle \left[{C}_{H}\right]\rangle\) or \(\langle \left[{C}_{V}\right]\rangle\)) by relating them to the full 3 × 3 scattering coherency matrix; as follows (Sugimoto et al., 2023):

$${k}_{V}=\left[\begin{array}{l}{S}_{VV}\\ {S}_{VH}\end{array}\right]=\frac{1}{\sqrt{2}}\left[\begin{array}{ccc}1& -1& 0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{c}{S}_{HH}+{S}_{VV}\\ {S}_{HH}-{S}_{VV}\\ 2{S}_{HV}\end{array}\right]$$
(1)
$$\langle \left[{C}_{V}\right]\rangle = \langle {k}_{V}{k}_{V}^{+}\rangle =\left[\begin{array}{cc}\langle {\left|{S}_{VV}\right|}^{2}\rangle & \langle {S}_{VV}{S}_{VH}^{*}\rangle \\ \langle {S}_{VV}^{*}{S}_{VH}\rangle & \langle {\left|{S}_{VH}\right|}^{2}\rangle \end{array}\right]=\left[\begin{array}{cc}{C}_{11}& {C}_{12}\\ {C}_{21}& {C}_{22}\end{array}\right]$$
(2)
$${k}_{H}=\left[\begin{array}{l}{S}_{HH}\\ {S}_{HV}\end{array}\right]=\frac{1}{2}\left[\begin{array}{lll}1& 1& 0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{c}{S}_{HH}+{S}_{VV}\\ {S}_{HH}-{S}_{VV}\\ 2{S}_{HV}\end{array}\right]$$
(3)
$$\langle \left[{C}_{H}\right]\rangle =\langle {k}_{H}{k}_{H}^{+}\rangle =\left[\begin{array}{cc}\langle {\left|{S}_{HH}\right|}^{2}\rangle & \langle {S}_{HH}{S}_{HV}^{*}\rangle \\ \langle {S}_{HH}^{*}{S}_{HV}\rangle & \langle {\left|{S}_{HV}\right|}^{2}\rangle \end{array}\right]$$
(4)

Henceforth, we will denote \(\left[{C}_{H}\right]\) and \(\left[{C}_{V}\right]\), by the symbol \(\left[{C}_{2}\right]\).

Dual Polarization Model-Based Scattering Power Decomposition

The established relationships were employed to extend the scattering power decomposition using dual-polarization data. Accordingly, we developed a four-component scattering power model for dual-pol data to represent scattering in tropical forests: the volume scattering power, oriented dipole scattering power, helix scattering power, and surface scattering power as shown in Fig. 2. A covariance matrix \(\langle \left[{C}_{2}\right]\rangle\) obtained in dual-polarization (\({S}_{VV}\) and \({S}_{VH}\) or \({S}_{HH}\) and \({S}_{HV}\)) mode can be expanded into four scattering matrices as:

Fig. 2
figure 2

Implementation flowchart of the proposed 4SD algorithm

Full size image
$$\begin{array}{c}\langle \left[{C}_{2}\right]\rangle ={P}_{g}{\langle \left[{C}_{2}\right]\rangle }_{g}+{P}_{v}{\langle \left[{C}_{2}\right]\rangle }_{v}+{P}_{h}{\langle \left[{C}_{2}\right]\rangle }_{h}+{P}_{od}{\langle \left[{C}_{2}\right]\rangle }_{od}\end{array}$$
(5)

In Eq. (5), \({\langle \left[{C}_{2}\right]\rangle }_{g}\),\({\langle \left[{C}_{2}\right]\rangle }_{v}\), \({\langle \left[{C}_{2}\right]\rangle }_{h}\) and \({\langle \left[{C}_{2}\right]\rangle }_{od}\) are ground, volume, helix, and oriented dipole scattering power matrices, respectively. \({{P}_{g}, P}_{v}, {P}_{h}\) and \({P}_{od}\) are dual polarization-based ground, volume, helix and oriented dipole scattering powers, respectively.

The ground scattering power matrix can be expressed as

$$\begin{array}{c}{\langle \left[{C}_{2}\right]\rangle }_{g}=\left[\begin{array}{ll}1& 0\\ 0& 0\end{array}\right]\end{array}$$
(6)

The covariance matrix for the volume scattering power is adopted by considering the uniform distribution of dipole structures in the vegetation area, as follows:

$${\langle \left[{C}_{2}\right]\rangle }_{v}=\frac{1}{4}\left[\begin{array}{ll}3& 0\\ 0& 1\end{array}\right]$$
(7)

The helix scattering model is able to account for imaginary part of the off-diagonal term in \(\langle \left[{C}_{V}\right]\rangle\) or \(\langle \left[{C}_{H}\right]\rangle\) (i.e., \({S}_{VV}{S}_{VH}^{*}\ne 0\) or \({S}_{HH}{S}_{HV}^{*}\ne 0\)) [18]. Helix scattering can be produced by compound targets (Singh et al., 2013), which generate circular polarization from all linear polarizations.

$${\langle \left[{C}_{2}\right]\rangle }_{h}=\frac{1}{2}\left[\begin{array}{cc}1& \pm j\\ \mp j& 1\end{array}\right]$$
(8)

When multiple targets exist within the same radar range, an oriented dipole matrix can be created through compound scattering phenomena. In the case of dual polarization, the oriented dipole scattering model accounts for the real part of the off-diagonal term and can be depicted as

$${\langle \left[{C}_{2}\right]\rangle }_{od}=\frac{1}{2}\left[\begin{array}{cc}1& \pm 1\\ \pm 1& 1\end{array}\right]$$
(9)

From Eq. (5–9), we can rewrite Eq. (5) as

$$\langle \left[{C}_{2}\right]\rangle ={P}_{g}\left[\begin{array}{cc}1& 0\\ 0& 0\end{array}\right]+\frac{{P}_{v}}{4}\left[\begin{array}{cc}3& 0\\ 0& 1\end{array}\right]+\frac{{P}_{h}}{2}\left[\begin{array}{cc}1& \pm j\\ \mp j& 1\end{array}\right]+\frac{{P}_{od}}{2}\left[\begin{array}{cc}1& \pm 1\\ \pm 1& 1\end{array}\right]$$
(10)

From (10), \({P}_{g},{P}_{v}\), \({P}_{od}\) and \({P}_{h}\) can be computed as

$$\pm {P}_{od}^{V}\pm j{P}_{h}^{V}=2\text{ Re}\langle {S}_{VV}{S}_{VH}^{*}\rangle +\text{j }2\text{Im}\langle {S}_{VV}{S}_{VH}^{*}\rangle$$
$${P}_{od}^{V}=2\left|Re\langle {S}_{VV}{S}_{VH}^{*}\rangle \right|$$
(11)
$$P_{h}^{V} = 2\left| {ImS_{VV} S_{VH}^{*} } \right|$$
$${P}_{v}^{V}=2\langle {\left|{S}_{VH}\right|}^{2}\rangle -{P}_{h}^{V}-{P}_{od}^{V}$$
(12)
$${P}_{g}^{V}= \langle {\left|{S}_{VV}\right|}^{2}\rangle -\frac{3}{4}{P}_{v}^{V}-\frac{1}{2}{P}_{od}^{v}-\frac{1}{2}{P}_{h}^{V}$$
(13)

Based on Eqs. (5–10), the algorithm is constructed, which is applied to calculate the 4SD scattering powers namely ground (\({P}_{g}\)), helix (\({P}_{h}\)), volume (\({P}_{v}\)), and oriented dipole (\({P}_{od}\)). The flowchart of the DP-4SD algorithm is illustrated in Fig. 2 for implementing on dual polarimetric SAR data.

Classification

Various classification techniques, such as Maximum Likelihood, Support Vector Machine, and Knowledge-Based Decision Trees, have been extensively studied and applied (Maghsoudi et al., 2013; Mishra et al., 2019; Varghese et al., 2016). However, the random forest (RF) classifier has proved efficient for the classification of vegetation using SAR datasets in several studies (Du et al., 2015; Waske & Braun, 2009). Therefore, the supervised Random Forest classifier was employed on the scattering power image generated from 3 and 4 SD to evaluate the accuracy of both scattering power decomposition techniques in identifying and classifying different LULC classes. Comprehensive technical details and discussions on the use of RF in remote sensing can be found in (Breiman, 2001).

The Random Forest classifier is of ensemble machine learning, which incorporates numerous decision trees to have a more accurate and robust output. The RF classifier builds a "forest" of decision trees, training each tree on a random selection of the training data using a randomly picked set of features (Belgiu & Drăguţ, 2016). Random Forests classification Eq. (14) introduced by (Breiman, 2001) is given below, which describes the process of constructing multiple decision trees and using them for classification by aggregating the outputs of all trees through majority voting.

$${\text{y}}^{ \wedge } = {\text{ mode }}\left( {{\text{T}}_{1} \left( {\text{x}} \right),{\text{ T}}_{2} \left( {\text{x}} \right),{\mkern 1mu} \ldots \ldots \ldots \ldots \ldots \ldots \ldots {\text{ Tn}}\left( {\text{x}} \right)} \right)$$
(14)

Where ŷ is the predicted class label for the input x.

T1(x), T2(x), …, Tn(x) are the predictions from the individual decision trees in the forest.

n is the number of trees in the forest.

Mode (·) denotes the majority voting among the predictions of all trees.

This study utilizes two reference data sets, a training set for classification purposes, and a testing set for post-classification validation for both sites. Both reference data sets were obtained using a random sampling method.

Results and Discussions

This study used the dual polarimetric L-band SAR datasets acquired over the San Francisco and Haldwani Forest regions from both the Advanced Land Observation Satellite (ALOS/PALSAR) and Advanced Land Observation Satellite-2 (ALOS-2/PALSAR-2) for a comparative analysis between the proposed DP-4SD and the existing DP-3SD model. Each color represents specific scattering powers represented qualitatively: blue represents ground scattering, red represents oriented dipole scattering, orange represents helix scattering, and green represents volume scattering. These scattering powers were combined in RGB color-coded images to assist image interpretation, target classification, and identification. This visualization technique detects differences in scattering mechanisms, thereby enhancing the SAR data analysis.

Quantitative Analysis of the Scattering Powers of 3SD and 4SD

Four regions of interest (RoIs) in RGB images of 4SD and 3SD, as shown in Fig. 3 were selected based on different forest compositions in the Haldwani region: teak, eucalyptus, mixed forest, and non-forested (NF) areas. The scattering power statistics of both the existing 3-component scattering decomposition algorithm and the proposed 4-component scattering decomposition algorithm were estimated using selected area classes to observe the differences in the scattering mechanisms between the two algorithms.

Fig. 3
figure 3

RGB images of a 4SD b 3SD for Haldwani with RoIs

Full size image

While the 3SD model has only three components of scattering powers that lead to the overestimation of volume scattering, the 4SD model has a new scattering component of oriented dipole scattering to reduce this overestimation. The most important observation from the investigation in the Haldwani region is the consistent decrease in volume scattering in 4SD compared to that in 3SD in all types of forests. This decrease signifies a better decomposition of the scattering mechanisms, especially for the forested classes. Volume scattering in 4SD decreased from 23.4 to 19.7% in teak forest, from 12.3 to 8.6% in eucalyptus forest, from 53.3 to 47.8% in mixed forest, and from 11.6 to 5.4% in non-forested areas, as shown in Fig. 4. To compensate for the reduced volume scattering, the oriented dipole scattering component contribution increased for all forest classes because the oriented dipole was generated by vegetation areas where tree branches were structured such that multiple scattering interactions of the radar signal occurred.

Fig. 4
figure 4

Model-based decompositions (3SD, 4SD) power statistics for Haldwani

Full size image

Thus, it was shown that in forest types with complex canopy structures, such as radar, signals incident to objects backscattered not only from the surface but also from the branches, resulting in compound scattering. Furthermore, this result showed that the dominance of surface scattering over low and sparsely vegetated land cover areas was observed in the case of forest plantations. This implies that the 4SD model captures the variability within the scattering behavior for various forest types. This signifies that the 4SD model is better suited for land cover classification and forest monitoring because of its improved accuracy.

These differences were more obvious in the case of San Francisco when comparing the 3SD and 4SD scattering decomposition models, (Fig. 5), underlining the improvement in interpretation arising from the inclusion of an oriented dipole scattering component in the 4SD model for complex urban and natural landscapes. The 4SD model discriminated between volume scattering and oriented dipole scattering, enabling better discrimination between urban structures, water body features, and vegetation class.

Fig. 5
figure 5

RGB images of a 4SD b 3SD for San Francisco with RoIs

Full size image

It improves the LULC classification accuracy, especially for areas that depict complicated scattering mechanisms, such as the oriented urban area and vegetation in San Francisco. This was demonstrated by a consistent volume scattering reduction when using the 4SD model compared to the 3SD model. This can only be attributed to the fact that the oriented dipole scattering component captures scattering in a more effective way, especially in urban landscapes where buildings and other structures are oriented such that the radar signal interacts with several surfaces to produce compound scattering effects.

This new scattering component provided a more detailed view of how different surfaces backscatter radar signals, thereby increasing the accuracy of the LULC. For instance, there was a significant reduction in ground scattering in the ground and urban patches obtained using the 4SD model. For instance, the ground scattering in the ground patch reduced from 90.3 to 77.8%, while that in the urban patch reduced from 97.5 to 84.8% (Fig. 6). This reduction is compensated for by the increase in the oriented dipole scattering component, which suggests that urban features, such as small structures and grid-like street layouts, are oriented in a way that contributes to compound scattering, which aligns along the radar’s line of sight, giving rise to scattering that cannot be modeled by the 3SD model.

Fig. 6
figure 6

Model-based decompositions (3SD, 4SD) power statistics for San Francisco

Full size image

The 4SD model has the ability to distinguish surface and small urban structures and identify the specific scattering behavior of streets and buildings, proving that this decomposition algorithm outperforms the existing algorithms. It can be analytically seen from the given that water bodies return a high percentage of ground scattering of approximately 90%, which is represented by a blue color in RGB visualizations. A strong ground-scattering signature is expected because the smooth surface of the water reflects most of the radar signal back to the sensor. Again, it is this hilly topography with slopes of varying gradients from gentle to steep, which again illustrates the 4SD model advantage. In vegetated patches, the orientation of the slopes with respect to the radar’s line of sight forms an oriented surface configuration such that cross-polarization effects appear.

This slope-induced cross-polarization led to an increase in volume scattering. Slopes in low-and sparsely vegetated areas also resulted in an enhanced ground-scattering contribution due to slopes, as observed by the 4SD model. Such terrains have a better representation of the scattering power owing to their ability to account for slope effects and the orientation of surface features in the 4SD model. Oriented dipole scattering becomes rather important in vegetated areas, where the structure of trees has an oriented structure of their branches, in such a way that compound scattering of the radar signal occurs. This occurs within some species where the structure of the branches can produce multiple scattering events that are captured more accurately by the 4SD model than by its predecessor, the 3SD model. It is this complex scattering behavior that the 4SD model can represent with better accuracy when trying to differentiate surface characteristics from densely vegetated, urban, and other land-cover types.

Comparison of LULC Classification Based on the Scattering Powers of 3SD and 4SD

The full potential of this DP-4SD model was investigated by applying it to land use/land cover classification and comparing it with classification based on the existing 3SD. A supervised random forest classification algorithm was employed, and the performance of the proposed 4SD model was further evaluated using data from dual polarization from two regions: Haldwani Forest in India and San Francisco urban region in the United States. These two sites were selected to test the model performance for different geographies and land cover classes. This testing was performed for two major purposes: checking the performance of the model in classifying areas containing a homogeneous class, such as a forested region in Haldwani, and its performance in heterogeneous LULC classes that are characteristic of highly complex urban environments, such as in San Francisco.

The 4SD model performed well, with a high overall accuracy of 85.69% and a kappa coefficient of 0.81 in the Haldwani Forest (Table 2), reflecting the strong agreement between the predicted and actual classifications. The outputs produced outperformed the existing 3SD model, achieving an overall accuracy of 84.90%. Classification maps are shown in Fig. 7.

Table 2 Haldwani
Full size table
Fig. 7
figure 7

Forest species classification based on a 4SD b 3SD

Full size image

The slightly improved accuracy of the 4SD model is considerably relevant, particularly in a complex ecosystem such as a forest, owing to the multiple above-mentioned scattering mechanisms present owing to the varied structure of trees, vegetation, and forest composition.

This is because the oriented dipole scattering component is included in the 4SD model, which significantly enhances the capability of this model to capture complex scattering behaviors depicting different forest species. Such an improvement in classification is important for forest cover mapping because it can provide more detailed differentiation between forest species, which may be important for applications in biodiversity assessments, carbon stock estimation, and forest management practices. It was also observed that the number of unclassified samples was reduced compared with the 3SD-based classification, further justifying the better performance of the 4SD model in the classification of forest types.

In the case of San Francisco, the DP-4SD model was highly effective, with an overall classification accuracy of 93.66%, in contrast to the existing 3SD-based classification model, which had an accuracy of 91.04% (Table 3). The high accuracy in a complex urban setting characterized by diverse built-up, vegetation, and water body classes underlines the increased ability of the 4SD model to classify different land-use classes in the urban landscape. Classification maps are shown in Fig. 8.

Table 3 San Francisco
Full size table
Fig. 8
figure 8

LULC classification based on a 4SD b 3SD

Full size image

Since DP-4SD relies only on HH and HV polarization channels, its ability to accurately distinguish urban areas based on double-bounce scattering is compromised. In contrast, full polarimetric data, which also includes VV and HV channels, provides more comprehensive information about the interactions between radar signals and urban surfaces, leading to better urban area classification. Urban areas exhibit more complex scattering behaviors, often requiring the inclusion of additional polarimetric channels (VV) to fully capture interactions such as the corner reflector effect, which happens when radar waves bounce between two perpendicular surfaces (e.g., buildings and streets). In dual-pol data (HH and HV), some of these intricate scattering patterns may not be fully resolved, and the DP-4SD model may overestimate surface scattering in urban areas or misclassify urban structures as natural features. For accurate urban area separation, other scattering decomposition techniques designed for full-pol data, such as the Freeman-Durden decomposition (Freeman & Durden, 1998) or Yamaguchi decomposition (Yamaguchi et al., 2005), are typically used. These techniques provide better differentiation of urban features by leveraging the full polarimetric channels to capture double-bounce scattering. In the absence of quad-pol data, dual-pol approaches like DP-4SD can be applied for urban classification.

Changes in the area covered by forest plantation species, according to 3SD and 4SD, are evident in Fig. 9. The model developed based on 4SD was finer in the demarcation of forest species, hence more accurate classification and representation of actual forest cover, which may have implications for the improved monitoring of deforestation and forest health.

Fig. 9
figure 9

Area (km2) covered by different LULC class

Full size image

The DP- 4SD model significantly reduced the number of unclassified samples to 0.701% in the DP- 4SD model-based classification compared with 2.88% in the classification based on the 3SD model. This marked reduction in unclassified regions ensures less presence of unclassified areas in the LULC classification.

This is important for land cover mapping, where unclassified areas can lead to gaps in the analysis and misinformed decision making. A comparison of the areas covered by different LULC classes based on the 3SD and 4SD models revealed that in the 4SD based classification there was a reduction in the vegetation class area by approximately 6 km2, which was covered by an increase in the urban class area (Fig. 9).

This can be explained by the fact that the oriented dipole scattering component within the 4SD model improves the characterization of the scattering behavior of urban structures and oriented branches within vegetation, decreasing the misclassification of urban areas into vegetated areas. The model DP-4SD has enhanced the detection of subtle variations in surface characteristics, vegetation structures, and building features by fully exploiting dual-polarization SAR data. Greater interpretation of polarimetric signals with enhancement in LULC classification may be helpful for better environmental monitoring and resource management, along with urban development. The implications of the DP-4SD model are multi-faceted. With the 4SD model, forest monitoring for changes in forest cover and early signs of deforestation or degradation can be detected to monitor health and structure using time-series dual-PolSAR data. This would result in a more timely and efficient approach for conservation, informed management of forests, and better understanding of shifts in forest dynamics due to climate change and human interference. The 4SD model can improve the classification of diverse urban structures and enhance the analysis and management in cities. Classification maps detailing the precise distribution of urban green spaces, buildings, and other classes of land use will be of great assistance to urban planners and policymakers. Thus, improvement in classification performance will serve to improve policies towards urban growth, development of green infrastructure, and mitigation of the urban heat island effect. Moreover, the possibility of the model DP-4SD to provide a more accurate classification with fewer unclassified samples could be particularly useful in dynamic environments, such as those with rapid urban growth or in areas prone to natural disasters.

Conclusions

This study introduces a novel model-based four-component scattering power decomposition for dual-polarization synthetic aperture radar (dual-pol) data by developing an oriented dipole, a novel scattering component for DP-4SD. The inclusion of this new component not only improves the accuracy of the PolSAR decomposition compared to earlier models but also emphasizes the benefits of oriented dipole scattering power, especially in densely vegetated regions and oriented urban structures. The urban-oriented structures and structures of the vegetation that generated dipole and dihedral scattering were identified using the DP-4SD model. The DP-4SD model demonstrated significantly improved classification accuracy compared to existing methods across different LULC classes. The DP-4SD model has the potential to advance the understanding of both natural and urban environments through more accurate classification and interpretation of land-cover types.