Monitoring and Analysis of Coastal Salt Pans Using Multi-Feature Fusion of Satellite Imagery: A Case Study Along the Laizhou Bay

Abstract

Coastal ecosystems, located at the interface of terrestrial and marine environments, provide significant ecological functions and resource value. Coastal salt pans, as critical coastal resources with significant implications for coastal ecosystem health and resource management, have attracted extensive research attention. However, current studies on the extraction of spatiotemporal patterns of coastal salt pans remain relatively limited and superficial. This study takes coastal salt pans in Laizhou Bay as a case study, proposing a hierarchical classification method—Salt Pan Feature-Enhanced Fusion Image Random Forest (SPFEFI-RF)—based on multi-index synergy guidance and deep-shallow feature fusion, achieving high-precision extraction of coastal salt pans. First, a Modified Water Index (MWI) and Salt Pan Crystallization Index (SCI) were constructed from image spectral features, specifically targeting the extraction of evaporation ponds. Concurrently, a salt pan sample dataset was developed for the DeepLabv3+ (DL) method to extract deep semantic features and perform multi-scale feature fusion. Subsequently, a three-channel fusion strategy—R(MWI)-G(SCI)-B(DL)—was employed to produce the Salt Pan Feature-Enhanced Fusion Image (SPFEFI), enhancing distinctions between salt pans and background land cover. Finally, the Random Forest (RF) classifier using shallow spectral features was applied to extract salt pan information, further optimized by spatial domain denoising techniques. Results indicate that the SPFEFI-RF approach effectively extracts coastal salt pan features, achieving an overall accuracy of 92.29% and a spatial consistency of 85.14% with ground-truth data. The SPFEFI-RF method provides advanced technical support for high-precision extraction of global coastal salt pan spatiotemporal characteristics, optimizing coastal zone management decisions and promoting the sustainable development of coastal ecosystems and resources.

1. Introduction

Coastal ecosystems, located at the terrestrial–marine interface, nurture diverse and unique habitats such as mangroves, coastal wetlands, and salt pans, playing an irreplaceable role in providing ecosystem services and supporting socio-economic development in coastal regions [1].

Coastal salt pans are an important resource and land use type in shoreline regions worldwide. As widely distributed production spaces and ecological units, they not only embody a long history of salt making but also play key roles in maintaining regional biodiversity, regulating local climate, and sustaining coastal ecological balance. However, under climate change, rising sea levels and intensified human activities, salt pan ecosystems face mounting pressure of degradation, conversion, and fragmentation. Their sustainable management has, therefore, become a transboundary challenge. Although salt pans are globally extensive and highly heterogeneous, and remote sensing has been applied to identify and monitor salt pans in local areas, there is no generalizable extraction method that operates across climatic zones and landforms and accommodates diverse spectral behaviors. This gap prevents the global assessment and coordinated governance of salt pan resources.

Consequently, high-accuracy extraction of coastal salt pans has become a research focus in recent years. Sridhar highlighted the importance of water body extraction in coastal zones and, based on the reflectance properties of water, proposed the Salt Pan Index (SpI) to jointly extract salt pans and aquaculture ponds [2]. Building on this line of work, Jin [3], Chen [4], and Ni [5] combined spectral and textural features—such as the Laplacian operator [4], Hessian matrix [4,5] and local spatial similarity [6] (Appendix A.3)—to derive salt pan targets. Machine learning approaches have also offered new technical routes: Safaee [7] and Ji [8] achieved high precision extraction using learning-based classifiers. Nevertheless, because salt pans, aquaculture ponds, and tidal flats often exhibit highly similar spectral and spatial signatures [6], methods relying on spectra alone struggle to separate classes; texture-based methods require careful feature design and are sensitive to window size and orientation, which limits model generalization.

Given these limitations in remote sensing feature extraction methods, recent research has increasingly adopted multi-feature extraction (including spectral, textural, shape-based, indices [9], and automated deep learning-derived features [10]) and information fusion strategies (such as multi-source [11], multi-temporal [12], and feature/decision-level fusion [13]). By combining multi-dimensional features and complementary information, these approaches significantly improve classification accuracy and model robustness. Nevertheless, for the specific case of salt pan extraction, challenges such as spectral confusion, sensitivity to textural parameters, and limited generalization capabilities remain, and more effective solutions are urgently required.

In particular, current methods struggle to effectively distinguish between evaporation ponds and aquaculture ponds within salt pans. Efficient extraction and integration of their unique spectral and spatial response characteristics, as well as the construction of models robust enough to accommodate complex environmental variations, remain open challenges.

To address the persistent confusion between evaporation ponds and aquaculture ponds, this study proposes a multi-feature collaborative extraction framework that establishes a new “spectral–semantic–machine-learning” paradigm for salt pan mapping. Because different water bodies have similar spectral responses, spectral indices alone cannot reliably distinguish evaporation ponds from other water samples; therefore, to capture complete salt pan characteristics, we extracted features for all samples of water containing evaporation ponds as well as crystallization ponds. We first design two targeted indices—the Modified Water Index (MWI) and the Salt Pan Crystallization Index (SCI)—to emphasize evaporation and crystallization features. To suppress residual non-salt pan water, we introduce, for the first time in salt pan extraction, the encoder–decoder Deeplabv3+ (DL) semantic segmentation model to resolve water-body details, particularly separating aquaculture ponds from evaporation ponds. We then implement an R(MWI)-G(SCI)-B(DL) three-channel fusion strategy to merge these outputs and produce a Salt Pan Feature-Enhanced Fusion Image (SPFEFI), which markedly improves separability from the background. Finally, we apply a Random Forest (RF) classifier to SPFEFI for coastal salt pan mapping and evaluate the effectiveness of this approach using a representative salt pan region in the Laizhou Bay coastal zone. Research on salt pan extraction provides essential evidence for understanding the coupling between resource use and ecological processes in coastal zones. These findings offer practical guidance for sustainable coastal management and balancing regional economic development with environmental protection.

2. Materials and Methods

2.1. Study Area Overview

The Laizhou Bay coastal zone lies along the north-eastern coast of Shandong Province, China, spanning approximately 118°30′–120° E and 37°00′–37°40′ N (Figure 1). The bay covers about 7000 km², the mainland shoreline is ~319 km long, and the mean water depth is ~10 m; overall relief is gentle. The climate is a typical temperate monsoon regime with four distinct seasons. Mean annual temperature is ~12.5 °C, and mean annual precipitation is ~600 mm, distributed very unevenly through the year and concentrated in July–September; potential evapotranspiration greatly exceeds precipitation.

Laizhou Bay is characterized by extensive, contiguous muddy tidal flats with exceptionally rich intertidal resources. The flats may extend several kilometers seaward with very low gradients, forming distinctive wetland systems. The Laizhou Bay coastal zone encompasses diverse land-cover types, including tidal flats, wetlands, salt marshes, mangroves, and coastal water, as well as varied land use categories such as agriculture, fisheries, salt production, industry, and tourism. Such abundant natural resources, combined with diversified land utilization patterns, position this area as critically significant for both regional economic development and ecological conservation. Laizhou Bay represents a typical coastal zone region, where salt extraction is primarily conducted through the evaporation of pumped subterranean brine [14].

2.2. Dataset

We used Landsat-8 OLI/TIRS imagery. Using Google Earth Engine (GEE), we retrieved scenes and performed radiometric calibration and atmospheric correction. Overall, cloud cover was low; the study area itself was cloud-free, and subsequent time-series processing mitigated any residual cloud effects. Data screening criteria are summarized in

Table 1. Image information.

No.	Date	Path/Row	Cloud Cover	Spatial Resolution (m)
1	18 April 2021	121/034	15.00%	30
2	5 June 2021	121/034	14.74%	30
3	21 June 2021	121/034	26.95%	30

. We employed visible and infrared band combinations at a 30 m spatial resolution.

2.3. Methods

In this section, we first selected three images with distinct salt pan features from 2021 and performed multi-temporal processing to generate the 2021meanImage. Subsequently, to meet research requirements, we extracted the land portion of the Laizhou Bay coastal zone by deriving the coastline from the 2021meanImage and extending it 20 km inland to define the study area (Figure 1). Following this, feature extraction was conducted for the MWI, SCI, and DL indices, which were then fused using an R(MWI)-G(SCI)-B(DL) composite to generate the SPFEFI. Finally, the SPFEFI was processed using the Random Forest algorithm (Figure 2).

2.3.1. Remote Sensing Indices

• Modified Water Index

To effectively enhance the extraction accuracy of evaporation ponds in salt pans and improve spectral separability between water and non-water bodies, this study proposes a Modified Water Index (MWI). Based on a comprehensive analysis of existing water indices—the Normalized Difference Water Index (NDWI) [15] and the Modified Normalized Difference Water Index (MNDWI) [16]—the spectral differences between water and non-water bodies were analyzed in detail (Appendix A.1 MWI). The MWI employs blue (${\rho }_{Blue}$), red (${\rho }_{Red}$), and short-wave infrared (${\rho }_{SWIR1}$) bands. These formulas are given as follows:

$$MWI={\displaystyle \frac{{\rho }_{SWIR1}-{\rho }_{Red}-{\rho }_{Blue}}{{\rho }_{SWIR1}+{\rho }_{Red}+{\rho }_{Blue}}}$$

(1)

$$NDWI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{NIR}}{{\rho }_{Green}+{\rho }_{NIR}}}$$

(2)

$$MNDWI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{SWIR}}{{\rho }_{Green}-{\rho }_{SWIR}}}$$

(3)

2.

Salt Pan Crystallization Index

Existing salinity indices, such as the Salinity Index (SpI) and Salinity Sensitivity Index (SSI), are primarily developed for monitoring soil or surface salinization. However, these indices struggle to accurately identify crystallization pools in salt pans. To effectively extract crystallization ponds and improve their spectral distinguishability from background land features, this study proposes a novel Salt Pan Crystallization Index (SCI). Based on detailed spectral analyses of salt pans and typical interference features (Appendix A.2 SCI), SCI optimizes band combinations to enhance extraction accuracy. The specific formulas are as follows:

$$SpI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{SWIR}}{{\rho }_{Green}+{\rho }_{SWIR}}}$$

(4)

$$SSI={\displaystyle \frac{{\rho }_{blue}}{{\rho }_{red}}}$$

(5)

$$SCI=\max \left\{({\rho }_{red}-{\rho }_{blue})-0.3\left[({\rho }_{NIR}-{\rho }_{SWIR2})-({\rho }_{red}-{\rho }_{SWIR2})\right]\right\}{\rho }_{blue}$$

(6)

2.3.2. DeepLabv3+

• Network Overview

DeepLabv3+ is a semantic segmentation model proposed by Google based on an encoder–decoder architecture (Figure 3). Its core innovations lie in combining multi-scale feature extraction and detail recovery mechanisms [17]. The model employs MobileNetV2 as its backbone, extracting features after two and four downsampling operations as foundational feature layers [18]. During encoding, multi-scale context information is captured through an Atrous Spatial Pyramid Pooling (ASPP) module, producing high-level semantic features upon fusion [19,20]. During decoding, the encoder outputs undergo a four-fold upsampling and are aligned and stacked spatially with shallow-layer features. Effective fusion of deep semantic information and shallow details is achieved via depthwise separable convolutions, generating high-precision segmentation results [21,22].

2.

Dataset Construction

On the GEE platform, we assembled a multi-temporal Landsat dataset by selecting summer scenes (June–September) from 2020 to 2024 with <30% cloud cover (

Table 2. Dataset information.

Cloud Coverage	Sensor	Date
<30%	LANDSAT 8 LANDSAT 9	May 2020~July 2020 May 2021~July 2021 May 2022~July 2022 May 2023~July 2023 May 2024~July 2024

). To address insufficient valid samples of salt pans, data augmentation strategies—including 90° rotations, mirroring, and adjustments to saturation, brightness, and contrast—were applied to increase dataset scale and diversity.

During the sample annotation phase, the images were divided into small patches to create semantic labels. Those containing cloud cover or unclear features of the salt pans were excluded. Representative examples of the semantic labels are shown in Figure 4. Semantic labels were manually annotated using visual interpretation, categorizing features into crystallization ponds, evaporation ponds, and background based on their spectral and textural characteristics, preserving spatial relationships and structural integrity among categories. The resulting semantic segmentation dataset contains 12,581 labeled samples, providing a robust foundation for training deep learning models.

2.3.3. Random Forest (RF)

Random Forest (RF) is a classification algorithm based on Bagging ensemble learning techniques (Figure 5), which significantly enhances classification stability and prediction accuracy by constructing multiple decision trees and aggregating their predictions [23]. The RF algorithm effectively handles high-dimensional data and mitigates multicollinearity issues [24,25]. The cooperative prediction and ensemble learning mechanisms of multiple decision trees substantially reduce the risk of overfitting [26,27]. Thus, this multi-feature collaborative method provides a comprehensive and reliable solution for extracting coastal salt pans.

2.4. Accuracy Assessment Method

This study employs a Kappa coefficient, Overall Accuracy, Precision, and F1-score as accuracy evaluation metrics (

Table 3. Accuracy metric formulas.

Accuracy	Formulation
Kappa	${\displaystyle \frac{Precision - \frac{(TP + FP) \times (TP + FN) \times (FP + TN)}{{(TN + TP + FP + FN)}^2}}{1 - \frac{(TP + FP) \times (TP + FN) \times (FP + TN)}{{(TN + TP + FP + FN)}^2}}}$
Overall Accuracy	${\displaystyle \frac{TP}{TP+FP}}$
Precision	${\displaystyle \frac{TP}{TP+FN}}$
F1 score	${\displaystyle \frac{UA\times PA}{UA+PA}}\times 2$

). True Positive (TP) denotes pixels correctly classified into a specific class. True Positives (TP) represent the number of samples that truly belong to a class and are correctly classified. False Positives (FP) denote pixels incorrectly classified into a class.

3. Results

3.1. Multi-Temporal Image Fusion

In salt production, environmental factors dynamically regulate the states of various pond sections, causing significant temporal changes in evaporation and crystallization ponds throughout the harvesting cycle, leading to variations in surface characteristics and spectral reflectance. Multi-temporal image fusion optimally preserves real land-cover characteristics of salt pans. Using three images from spring and summer of 2021 as an example, image fusion effectively mitigated the influence of dynamic variations during the salt production cycle (Figure 6).

In the temporal analysis of remote sensing imagery within the salt pan region, the dynamic variation characteristics of crystallization and evaporation ponds are distinctly observable. For the crystallization pond area (highlighted by the yellow ellipse in the first row), significant temporal differences are evident across different image dates. The image from 18 April 2021 reveals that the crystallization ponds appear darker, indicating that some ponds had not reached the typical crystallization stage (Figure 6a). On 5 June 2021, the crystallization ponds collectively displayed a significantly darker tone, with limited spectral separability from the surrounding features (Figure 6b). By 21 June 2021, the distinction between crystallization ponds and their surroundings improved noticeably, and the spectral features of the ponds became more pronounced (Figure 6c). Therefore, to address the inconsistent characteristics of crystallization ponds across temporal images, a multi-temporal fusion approach was employed using the first three images (Figure 6d). The fused image shows a uniform, bright tone with clear spectral features, consistent with optimal crystallization conditions. Simultaneously, the fusion enhanced spatial uniformity within adjacent evaporation ponds.

For the evaporation ponds (highlighted by the yellow ellipse in the second row), optical features varied temporally. On 18 April 2021, no salt crystallization was observed in these ponds (Figure 6e). By 5 June 2021, minor salt crystallization began to appear (Figure 6f). By 21 June 2021, salt crystallization significantly increased (Figure 6g). The fused image retained the crystallization information from 21 June while better representing the typical optical characteristics of evaporation ponds (Figure 6h). Furthermore, compared with single-temporal images, the fused evaporation pond areas exhibited enhanced spatial consistency in tone.

3.2. Spectral Index Feature Extraction

We compared the results of SCI and MWI using the original images and conducted on-site observations at the salt field, confirming that the results of MWI and SCI extraction were consistent with the real surface information. Comparisons of results with existing indices are provided as follows.

3.2.1. Modified Water Index (MWI)

Based on the Modified Water Index (MWI) described in Section 2.3.1, water and non-water features were extracted and comparatively analyzed against traditional indices such as Normalized Difference Water Index (NDWI) and Modified NDWI (MNDWI). In the extraction of specialized water bodies such as salt pans, NDWI and MNDWI exhibited notable band limitations. As illustrated in Figure 7 (red ellipse region), NDWI fails to effectively differentiate spectral features between water bodies and salt pan crystallization ponds, resulting in highly similar extraction outcomes and misclassification of embankments between evaporating ponds as water bodies. Although MNDWI improves the distinction between evaporating ponds and crystallization ponds, it introduces abrupt tonal changes, significant spectral heterogeneity, and unstable extraction results.

In contrast, MWI not only achieves high extraction accuracy but also maintains spatial continuity, significantly improving spectral uniformity within evaporating pond areas. Additionally, while enhancing water information, MWI effectively distinguishes water bodies (including evaporating ponds) from crystallization ponds within the salt pan regions.

As demonstrated in Figure 7, the MWI, based on spectral features, exhibits clear advantages in distinguishing water and non-water areas. By optimizing the combination of visible and short-wave infrared bands, MWI significantly enhances spectral separability between crystallization and evaporating ponds and accurately identifies water bodies from background features. Moreover, MWI demonstrates robust performance against interference from complex backgrounds, significantly reducing false detections from surrounding dry surfaces and built-up features, thus enhancing the overall accuracy and reliability of salt pan extraction.

3.2.2. Salt Pan Crystallization Index

The Salt Pan Crystallization Index (SCI), introduced in Section 2.3.1, was applied for extracting crystallization ponds using a synergistic analysis of visible and infrared bands. Its performance was compared with existing salt indices, SpI and SSI, as shown in Figure 7.

Comparative analysis of SpI, SSI, and SCI, validated using false-color composites, indicates that traditional soil salinity indices (SpI and SSI) perform poorly in crystallization pond extraction. This inadequacy arises because these indices were primarily developed for soil salinization detection and fail to capture spectral characteristics of high-salinity water bodies, resulting in blurred boundaries and fragmented internal structures. Conversely, SCI enhances the distinctive high reflectance features specific to crystallization ponds while suppressing background interference, substantially reducing false detections from adjacent soils and vegetation, thus preserving pond boundary continuity. The SCI extraction closely matches the spatial distribution in the false-color composite, especially excelling in detecting lower-concentration crystallization ponds (yellow ellipse, Figure 7), where SpI and SSI exhibit varying degrees of omission.

By enhancing the spectral characteristics of crystallization ponds, SCI effectively mitigates the sensitivity of traditional indices to mixed pixels, clearly delineating crystallization ponds from other land cover types. The central contribution of SCI lies in its critical role within the hierarchical framework for salt pan feature extraction, constituting an indispensable element for accurately delineating crystallization pond structures.

3.3. Extraction Using DeepLabv3+

3.3.1. Model Construction

The semantic segmentation model was constructed based on the DeepLabv3+ framework, employing MobileNetV2—a lightweight neural network—as its backbone. To optimize model performance, a phased, progressive training strategy was implemented. In the initial phase, the MobileNetV2 pre-trained on ImageNet served as initial weights and was trained for 100 epochs. For the second phase, the model weights corresponding to the lowest validation loss from the first phase (epoch = 100, val_loss = 0.166) were used as the starting point for another 100 epochs of training. This iterative, pyramid-style training method gradually approaches the optimal solution, significantly enhancing the model’s capability for feature extraction (Figure 8).

3.3.2. Inference Results of Deeplabv3+

A multi-model selection strategy was employed during inference. Through predictive comparisons of the weight files generated after each epoch, visual assessment indicated that the weight file from epoch 105 delivered the optimal prediction performance. Six sets of model weights (epochs: 05, 50, 95, 105, 140, 195) corresponding to different stages of validation loss were selected for comparative analysis (Figure 9). Significant variations were observed between epochs regarding the misclassification of water bodies (indicated by yellow ellipses) and the missed detection of evaporating ponds in the salt pan (indicated by green ellipses). Among the evaluated models, the epoch-105 model (val_loss = 0.141) exhibited the best detection performance, presenting the fewest misclassifications of water bodies and uniquely avoiding any omission of evaporating ponds in the salt pan. Notably, subsequent epochs (e.g., 140 and 195) demonstrated lower validation loss and higher mIoU scores (Figure 10); however, these models exhibited pronounced missed detections during inference on the test dataset. Experimental results indicate that the optimal inference performance does not necessarily correspond to the models with the lowest validation loss or the highest mIoU scores. Instead, the epoch-105 model—trained just prior to overfitting—demonstrated enhanced robustness within the 2021 dataset, as evidenced by its accurate preservation of the geometric structural features of salt-pan regions across the different scenes used to create the summer composite. This finding validates the effectiveness of the phased fine-tuning strategy, further emphasizing that appropriately timed early stopping is crucial to preventing overfitting and consequent performance degradation.

3.4. Result of RGB Strategy Multi-Feature Fusion

The proposed Salt Pan Feature-Enhanced Fusion Image (SPFEFI) method collaboratively utilizes spectral indices and deep learning features to enhance the visual interpretability of salt pan surface characteristics. An RGB color space compositing strategy was employed, assigning independently derived grayscale images—MWI, SCI, and DL features—to the R, G, and B channels, respectively. This linear combination of grayscale values produces a composite color image that facilitates a multi-dimensional representation, effectively revealing spatial relationships among features through channel overlay (Figure 10).

Regions with high values in the red channel (R ≈ 255) represent all water bodies detected by MWI, including salt pan ponds. Highlighted regions in the green channel (G ≈ 255) correspond to crystallization ponds identified by SCI. High-intensity areas in the blue channel (B ≈ 255) represent the main salt pan areas classified by the DL model. Overlapping the three primary colors produces mixed-color areas indicating feature consistency: magenta regions (R + B) indicate areas identified by both MWI and DL models as water bodies, and yellowish-white areas (G + B) indicate spatial coupling between SCI and DL predictions. Other colors reflect land-cover types with lower salt pan-feature consistency; for instance, pure red areas represent aquaculture ponds or other water bodies. This imaging approach employs a color-semantic encoding scheme to enable intuitive visualization of multidimensional salt pan features, enhancing their distinction through collaborative feature interactions, and facilitating effective chromatic differentiation between various salt pan developmental stages, including evaporating and crystallization ponds.

Although the fusion of multiple features significantly enhances salt pan characteristics, inconsistencies occur in the SPFEFI results (Figure 10) due to intrinsic differences in the mechanisms through which MWI identifies water bodies and the DL model classifies salt pans.

In Figure 11a,c, the red anomaly within the yellow ellipse indicates evaporating ponds misclassified as water bodies. This occurs due to the DL model’s inability to accurately recognize evaporating ponds, resulting in the absence of a strong blue channel response (B < 255) and leaving only a dominant red signal from MWI (R ≈ 255). Although this red shade visually resembles pure aquaculture ponds, the slight non-zero green channel response introduces subtle salt pan characteristics, resulting in a darker, less saturated red tone that can be distinguished from the typical bright red color of aquaculture ponds.

The magenta-colored anomaly within the green ellipse in Figure 11a corresponds to aquaculture ponds mistakenly classified as salt pans due to the DL model incorrectly identifying the highly reflective aquaculture ponds as salt pans. This triggers a strong blue channel response (B ≈ 255), overlapping with the MWI red channel (R + B). However, the absence of SCI green channel responses (G ≈ 0), combined with spectral differences from actual evaporating ponds, causes these areas to display cooler-toned or unusually bright magenta shades, distinguishable from the warmer-toned magenta associated with genuine evaporating ponds that have slight green-channel signals.

The misrepresentation of evaporating pond features in Figure 11 was subsequently addressed using the Random Forest classification approach.

3.5. Extraction of Salt Pan Information

The method integrating SPFEFI fused image with the Random Forest (RF) classifier effectively extracted salt pan information along the Laizhou Bay coastline (Figure 12a). In localized analyses, the SPFEFI-RF method demonstrated superior performance (Figure 12b–d). Despite complex backgrounds comprising bare soil, vegetation, and water bodies surrounding the salt pans, boundaries and internal structural details of the salt pan targets were precisely distinguished and extracted (Figure 12b). Concerning spatial heterogeneity caused by highly similar spectral and textural characteristics between salt evaporation ponds and aquaculture ponds (indicated by the red ellipse in Figure 12c), the RF classifier accurately differentiated them by capturing subtle differences in their spectral feature space, thus preventing misclassification. Additionally, the method showed robust performance even in scenes containing disturbances from rivers, reservoirs, and bare soil. Although slight misclassification occurred in areas indicated by red ellipses (Figure 12d), this primarily resulted from inherent spectral and spatial similarities. Overall, the SPFEFI-RF method proved highly applicable to Laizhou Bay, achieving comprehensive and accurate extraction of large-scale, regularly distributed salt pan structures.

The RF classification method based on SPFEFI excels in accurately preserving detailed salt pan features, even against complex backgrounds. It effectively resolves the confusion between salt pans and spectrally similar land cover types, particularly aquaculture ponds. The minor local misclassification observed with reservoirs and bare soil suggests that future studies should consider incorporating finer spectral characteristics or spatial contextual information to further enhance the model’s ability to discriminate subtle differences.

3.6. Accuracy Assessment

Based on the accuracy evaluation approach described in Section 2.4, we quantitatively assessed the accuracy of the SPFEFI-RF method for salt pan extraction in the study area (

Table 4. Assessment of the SPFEFI RF method.

Accuracy	Crystallization Pond	Evaporation Pond
Overall Accuracy	92.29%	92.29%
Kappa	84.01%	84.01%
Precision	99.10%	83.06%
F1 score	90.34%	93.66%

All precision calculations were performed using the same sample set.

). The F1-scores obtained for crystallization ponds and evaporation ponds were 90.34% and 93.66%, respectively, and the overall classification accuracy reached 92.3%. These results demonstrate that the proposed method provides accurate and reliable extraction outcomes.

To further analyze the extraction accuracy and reliability of the SPFEFI-RF method, its results were compared with ground truth data and results obtained from the original RF classification approach (Figure 13). As illustrated by the multi-temporal fused image (Figure 13a) and a false-color image (Figure 13c), the experimental area exhibits complex land covers, with extensive water bodies and high spectral similarity among salt pans, other water bodies, and tidal flats, making classification challenging. The results obtained from RF classification using a multi-temporal fused image (Figure 13e) exhibited discontinuous salt pan boundaries, substantial internal omissions, poor spatial continuity, and widespread misclassifications throughout the study area. According to the analysis combined with Figure 13c, misclassified land cover types primarily included other water bodies, tidal flats, and built-up areas, reflecting the limited spectral discrimination capability of the original image. In contrast, incorporating both spectral and deep semantic features significantly enhanced the differentiation between salt pans and other land covers (Figure 13b). This resulted in clearer and more coherent salt pan boundaries, substantially improved internal continuity, and significantly reduced misclassifications, particularly in regions prone to confusion between water bodies and tidal flats (Figure 13e). Further comparison with ground truth data (Figure 13f) indicated that this feature-enhanced method markedly improved extraction accuracy compared with traditional RF classification, reduced fragmentation effects, and yielded an overall classification result superior to direct classification from the original image, achieving an overlap rate with ground truth data of 85.14%.

To more precisely evaluate the classification performance of the two image types, representative land cover categories within the study area were selected for localized comparative analysis (Figure 14). In the mixed zone of salt evaporation ponds and aquaculture ponds (red ellipse in Figure 14a), the classification result derived from multi-temporal images using the original RF approach (origin-RF) showed significant confusion, with aquaculture ponds frequently misclassified as salt evaporation ponds. In contrast, the SPFEFI-RF method effectively preserved the spatial structure of salt pan boundaries, with only a few isolated misclassified pixels, which were subsequently removed through post-processing in the RF classification. In the boundary area between crystallization ponds and bare soil (Figure 14b), both the multi-temporal fused image and SPFEFI demonstrated relatively accurate discrimination of bare soil. However, the SPFEFI-RF method produced more regularly aligned boundary pixels, exhibiting a higher spatial consistency with the ground truth contours. In the vegetation-covered area (red ellipse in Figure 14c), the origin-RF result exhibited evident salt-and-pepper noise, manifesting as scattered misclassification patches. By contrast, the SPFEFI-RF results showed a more natural transition between vegetated areas and salt pans, with significantly improved noise suppression. In summary, while the SPFEFI-RF method performed comparably to the multi-temporal image in differentiating bare soil, it yielded superior boundary delineation results. Furthermore, in areas with complex water bodies (Figure 14a) and vegetation edges (Figure 14c), the SPFEFI-RF method produced fewer misclassifications and closer alignment with the ground truth. These findings suggest that incorporating deep semantic features can further enhance the ability to distinguish spectrally similar land cover types.

4. Discussion

4.1. Factors Affecting the Extraction of Evaporation Ponds

In remote sensing imagery, evaporation ponds typically appear as densely clustered, grid-like structures, whereas a smaller number of evaporation ponds are irregularly shaped and loosely grouped. These features exhibit distinctive geometric characteristics. The main challenge in accurately extracting evaporation ponds lies in interference from spectrally and geometrically similar water bodies, such as aquaculture ponds. Effectively removing such interference is essential to ensuring both classification accuracy and the reliability of the application results. The proposed method effectively extracts salt pan evaporation ponds while successfully eliminating interference from unrelated water bodies (Figure 15).

Aquaculture ponds that share similar spatial arrangements with salt evaporation ponds (Figure 15a) pose a significant risk of misclassification due to their nearly identical spatial distribution patterns and spectral characteristics. Relying solely on spectral or semantic features is insufficient for effectively removing such interference; the integration of spatial-domain information is essential to achieving accurate differentiation. In contrast, aquaculture ponds arranged in narrow, linear strips exhibit clear textural distinctions from evaporation ponds (Figure 15b), although their spectral profiles still exhibit substantial overlaps. In this case, semantic features can effectively reduce or even eliminate such interference, and the boundary between evaporation ponds and aquaculture ponds is clearly defined in overlapping regions (Figure 15b). For spectrally similar yet spatially isolated water bodies such as reservoirs (Figure 15c), applying random forest classification to SPFEFI-enhanced imagery effectively removes these sources of interference and minimizes false positives.

4.2. Factors Affecting the Extraction of Crystallization Ponds

In remote sensing imagery, crystallization ponds in salt pans typically appear as regularly arranged rectangular or grid-shaped structures, characterized by well-defined boundaries and consistent geometric shapes. Their spectral characteristics vary significantly across space and time due to differences in salt concentration and crystallization stage. However, land cover types such as bare soil, buildings, alkali processing zones, and tidal flats exhibit spectral signatures similar to those of crystallization ponds. This spectral similarity poses significant challenges for the accurate interpretation of remote sensing imagery. Consequently, relying solely on spectral information for classification often results in misclassification. The proposed method effectively distinguishes crystallization ponds from spectrally similar land cover types, significantly improving the accuracy and reliability of remote sensing interpretation (Figure 16).

The proposed method demonstrates strong discriminatory capability when distinguishing crystallization ponds from spectrally similar surfaces such as bare soil (Figure 16a,b) and tidal flats (Figure 16c). In bare soil interference zones, the algorithm captures subtle spectral differences between soil and crystallization ponds, maintaining clear spatial separation and effectively suppressing misclassification caused by bare land (Figure 16a,b). In scenarios involving tidal flat interference, despite the high spectral similarity, the method maintains consistent extraction accuracy, indicating strong anti-interference robustness (Figure 16c). Even in cases where crystallization pond features are poorly expressed (Figure 16d), the method still accurately identifies the target class, ensuring both the completeness and reliability of the extracted results. This robustness stems from the algorithm’s optimized extraction of multidimensional salt pan features, enabling it to adapt effectively to varying scene complexities in crystallization pond detection.

4.3. Advantages and Limitations of the Proposed Method

This study presents a salt pan extraction method that integrates remote sensing indices, deep learning, and random forest classification. Although the method performs well overall, several limitations remain.

4.3.1. Advantages

First, the multi-feature collaborative strategy substantially improves separability between salt pans and confusing classes. The MWI and SCI indices enhance the spectral response of salt pans. In comparison with SCI, ASI [6] and NDSI [2] for crystallization-pond extraction, all three indices perform well, but SCI better distinguishes tidal flats from crystallization ponds. In addition, deep semantic features derived by Deeplabv3+ compensate for the limited use of spatial context in traditional spectral methods and remove several interferents that MWI and SCI alone cannot filter. Fusing these outputs into the RGB-composite SPFEFI further enlarges the contrast between salt pans and commonly confused classes such as tidal flats and aquaculture ponds, yielding higher classification accuracy than any single-feature approach. Secondly, the staged classification design balances efficiency and accuracy. The RF classifier performs a second-stage extraction on the feature-enhanced imagery, retaining the interpretability of shallow spectral cues while suppressing noise through spatial denoising. The final results achieved high accuracy (Kappa = 0.84) with fewer misclassifications among salt pan-confounding classes.

4.3.2. Limitations

Firstly, although the overall separability is high, land types such as bare soil and aquaculture ponds—due to their close similarity in both spectral and spatial domains—still occasionally cause misclassification. This issue arises mainly from the overlapping feature space in both spectral reflectance and semantic representation. Secondly, the methodological framework is relatively complex. Although the integration of MWI, SCI, DeepLabv3+, and RF enhances extraction accuracy, it also increases computational cost and necessitates substantial hyperparameter tuning. Finally, the performance of the model is highly dependent on training data. DeepLabv3+, in particular, is sensitive to the quantity and quality of annotated samples. In the absence of representative training data, the model’s generalization ability may degrade. Future work could explore the integration of higher-resolution temporal features or the incorporation of elevation data to further reduce false positives.

5. Conclusions

This study addressed the challenge of accurately extracting coastal salt pan features by developing a hierarchical classification method, SPFEFI-RF, which integrates multi-index guidance with deep–shallow feature fusion. The strength of this framework lies in its progressive design for salt pan extraction. Specifically, the MWI and the SCI were constructed based on spectral information from remote sensing imagery. A salt pan sample dataset was then established, and a deep learning model was employed to extract multi-scale deep semantic features. These outputs were further combined through an R(MWI)-G(SCI)-B(DL) three-channel fusion strategy to generate an SPFEFI, which significantly improved the separability of salt pans from background land covers. Finally, the RF classifier was applied to extract salt pan information from shallow features, and spatial denoising was performed to refine the final results.

By employing this feature-fusion strategy, salt pan information was effectively extracted, overcoming the limitations of traditional single-feature approaches and resolving the difficulty of distinguishing water bodies from evaporation ponds with highly similar spectral responses, particularly in separating evaporation ponds from aquaculture ponds. The overall accuracy of the extraction reached 92.29%, and the spatial consistency between the extracted salt pans and the ground truth was 85.14%. SPFEFI-RF demonstrated strong adaptability and robustness in complex coastal environments with multiple interfering land covers, achieving high-precision salt pan extraction. Salt pan mapping has profound implications for understanding resource use, ecological processes, and system stability in coastal zones. This research not only reveals the coupling mechanisms between human activities and ecosystems but also provides a scientific basis for optimizing land use, enhancing ecosystem resilience, and advancing sustainable development in coastal regions.

In future work, we plan to incorporate Digital Surface Model (DSM) data or Interferometric Synthetic Aperture Radar (InSAR) imagery as additional information sources. Integration of DSM data would enable the model to capture three-dimensional surface structures and elevation information, which is crucial for distinguishing land covers with similar spectral signatures but notable elevation differences, such as salt pans, water bodies, and flat terrain. Meanwhile, InSAR imagery can provide surface deformation data, facilitating the identification of internal structural characteristics and dynamic processes within salt pans. It is anticipated that multimodal data fusion will significantly enhance the segmentation accuracy and the model’s capability to interpret complex land-cover scenarios.