Fine-Grained Land Use Remote Sensing Mapping in Karst Mountain Areas Using Deep Learning with Geographical Zoning and Stratified Object Extraction

Abstract

Karst mountain areas, as complex geological systems formed by carbonate rock development, possess unique three-dimensional spatial structures and hydrogeological processes that fundamentally influence regional ecosystem evolution, land resource assessment, and sustainable development strategy formulation. In recent years, through the implementation of systematic ecological restoration projects, the ecological degradation of karst mountain areas in Southwest China has been significantly curbed. However, the research on the fine-grained land use mapping and quantitative characterization of spatial heterogeneity in karst mountain areas is still insufficient. This knowledge gap impedes scientific decision-making and precise policy formulation for regional ecological environment management. Hence, this paper proposes a novel methodology for land use mapping in karst mountain areas using very high resolution (VHR) remote sensing (RS) images. The innovation of this method lies in the introduction of strategies of geographical zoning and stratified object extraction. The former divides the complex mountain areas into manageable subregions to provide computational units and introduces a priori data for providing constraint boundaries, while the latter implements a processing mechanism with a deep learning (DL) of hierarchical semantic boundary-guided network (HBGNet) for different geographic objects of building, water, cropland, orchard, forest-grassland, and other land use features. Guanling and Zhenfeng counties in the Huajiang section of the Beipanjiang River Basin, China, are selected to conduct the experimental validation. The proposed method achieved notable accuracy metrics with an overall accuracy (OA) of 0.815 and a mean intersection over union (mIoU) of 0.688. Comparative analysis demonstrated the superior performance of advanced DL networks when augmented with priori knowledge in geographical zoning and stratified object extraction. The approach provides a robust mapping framework for generating fine-grained land use data in karst landscapes, which is beneficial for supporting academic research, governmental analysis, and related applications.

1. Introduction

Karst mountain areas, unique geological landscapes formed from soluble carbonate rocks like limestone and dolomite, cover about 25% of the world’s ice-free land area and supply drinking water to over 20% of the global population [1,2]. In China, with extensive carbonate rock distribution, karst landforms are well-developed and diverse, spanning a total of 3.444 million km², of which 0.907 million km² is exposed. The South China Karst region, centered in Guizhou, has an exposed area of 0.54 million km², making it the largest and most developed of the world’s three major karst regions [3]. However, due to the scarcity of acidic insoluble substances from parent rock weathering, soil formation is slow and soil layers are infertile in karst areas. Coupled with a climate featuring concurrent rainfall and heat, concentrated heavy rains, and ongoing human agricultural activities, this makes the region highly susceptible to severe soil erosion and accelerated karst rocky desertification [4,5]. The southwestern China karst region faces multiple challenges, including ecological fragility, pronounced human-land conflicts, and severe desertification, making it one of the key ecologically fragile areas in need of management [6,7]. Additionally, the region’s unique geographical conditions, such as cloudy and rainy weather and fragmented land features, result in a scarcity of high-quality satellite imagery data and severe pixel mixing [8]. Against the backdrop of the transition from traditional monitoring methods to intelligent cognition, there is an urgent need to establish a low-cost, refined, and quantitative regional monitoring and regulatory system.

Spatial units, fundamental for information organization and analysis, directly impact the soundness and effectiveness of spatial analysis. Traditional research often uses regular pixels from remote sensing (RS) datasets as analysis units. While this grid-based method is convenient and computationally efficient, it struggles to accurately delineate the fine boundaries and internal/external heterogeneity of geographical elements. As data technologies advance and geographical information analysis demands deepen, segmentation objects are increasingly replacing image pixels as basic analysis units in many studies and applications [9]. This shift not only enables precise definition of geographical element boundaries but also facilitates spatial and knowledge representation [10]. Moreover, using segmentation objects significantly reduces spatial data volume while increasing information content through their shape, internal texture, and spatial relationships. In karst mountain research, the delineation of fine-grained spatial objects is crucial. It can delineate vegetation coverage, human activity boundaries, and rock exposure units. This provides a more accurate reference spatial framework for regional ecological monitoring and resource management.

The advent of the geographical big data era has provided rich, high-resolution data in both space and time for fine-grained object segmentation. Consequently, growing research focuses on generating segmented objects [11,12,13] and supporting subsequent applications [14,15,16]. Object generation methods are diverse, including edge-based and region-based [17], supervised and unsupervised [18], traditional and deep learning approaches [19], as well as single-scale and multi-scale methods based on segmentation levels [20]. Edge-based methods represented by Robert, Prewitt, and watershed algorithms constitute an important component of traditional segmentation approaches. Their basic process can be summarized as detecting heterogeneous boundaries and filling interiors to determine objects, often requiring filtering and enhancement techniques to highlight edge features. However, these methods tend to produce problematic edges, including false, missing, shifted, and broken boundaries, affecting segmentation accuracy. Another classic category of traditional segmentation methods is region-based image segmentation, represented by K-means, Multi-Resolution Segmentation, and Mean-Shift algorithms. These methods operate on the principle of object internal feature similarity (color, spectral, texture, etc.), achieving segmentation through iterative merging or splitting of adjacent objects. The effectiveness of such methods heavily depends on threshold settings, and single-scale segmentation struggles to accommodate the non-uniform segmentation requirements across different geographical objects. Currently, segmentation methods are gradually being replaced by deep learning (DL) architectures, primarily because DL-based segmentation techniques can not only eliminate tedious manual processes through carefully designed feature extraction modules (such as convolutional layers, attention channels, and activation layers) but also demonstrate superior segmentation performance and potential semantic labeling capabilities [21,22,23].

Nevertheless, unlike common images, achieving good segmentation of remote sensing imagery remains highly challenging. Current data-driven deep learning models heavily rely on the quantity and quality of labeled data, while high-resolution remote sensing images exacerbate these issues due to their high intra-class variability and weak inter-class separability [24]. Some scholars have attempted to address data imbalance and class confusion problems through DL itself [24,25,26], while others have tried incorporating prior knowledge and constraint rules into the solution process [27,28,29]. The black-box nature of DL raises concerns about its reliability. To avoid significant errors, coupling advanced DL with rich prior knowledge becomes essential. Among diverse approaches, the strategy of geographical zoning and object stratification has been extensively introduced in complex remote sensing image segmentation, demonstrating significant effectiveness [29,30,31].

There are several recent studies focused on land mapping using remote sensing data in complex karst mountain areas with highly heterogeneous landscapes [32,33]. However, this applicability of combining the geographical zoning and object stratification strategy with DL in complex karst mountain areas remains unexplored. The existing methods do not fully consider the complexity of the Earth’s surface in karst mountain areas. This research gap is particularly significant given the distinctive characteristics of karst landscapes. It is expected to achieve a better mapping result with the support of prior knowledge, as the mapping of land use in karst mountain areas poses following two kinds of significant challenges. One is that karst mountain areas exhibit highly heterogeneous topography, resulting in land use objects with diverse morphological characteristics, including variable sizes, shapes, and spatial distribution patterns. The other is that different land use objects display distinct yet sometimes ambiguous spectral and textural signatures in very high resolution (VHR) RS imageries. These complex characteristics pose substantial challenges for accurate, fine-grained land use mapping in karst regions. How to create high-precision and detailed land use maps is currently a challenge in RS interpretation.

Based on this background, this paper proposes to achieve fine-grained land use mapping of complex karst mountains by focusing on the zonal and stratified strategy combined with the HBGNet model. Specific research contents include: (1) constructing regional prior knowledge via geographical zoning based on geomorphological features and existing data to extract key spatial constraint boundaries, (2) selecting DL segmentation networks adapted to complex terrain conditions and implementing hierarchical strategies in stratified land use object extraction to ensure model inference performance, and (3) preparing high-precision land use maps for the karst region. The study area was selected at the Huajiang section of the Beipanjiang River canyon at the junction of Guanling and Zhenfeng counties in Guizhou Province, China, which represents a typical karst plateau canyon region. The study area is dominated by moderate to severe rocky desertification, and human activities in managing rocky desertification have led to extensive land reclamation, creating an urgent need for precise monitoring of land use. The research will provide important data support for karst mountain ecological environment monitoring, resource management, and regional sustainable development.

2. Materials

2.1. Study Area

The study area, as shown in Figure 1, is located in the Huajiang section of the Beipan River Basin in Guizhou Province, encompassing Guanling County and Zhenfeng County (105.15–105.56°E, 25.07–26.05°N). Situated in the transitional slope zone between the Yunnan-Guizhou Plateau and the Guangxi Hills, this region is characterized by intensively developed karst landforms. The northwestern Guanling County features steep fault-block mountains, while the southeastern Zhenfeng County forms a stepped terrace structure. Together, these geomorphic units create a three-dimensional landscape system with vertical elevation differences exceeding 1600 m, incorporating multiple karst morphologies, including dissolution hill plains, peak cluster depressions, and karst canyons.

Under the long-term interplay between human activities and karst processes, the study area has developed a highly heterogeneous land use system. The topographic segmentation effects of deep-incised valleys and multi-level terraces have shaped a “fragmented-terraced” binary structure across the landscape. Specific land use patterns include concentrated woodland and orchards on terrace summits, rain-fed terraced croplands embedded along slopes, linear farming zones following river valleys, and scattered village settlements. This topography-constrained land use pattern results in highly irregular plot geometries and significant spectral confusion between different land categories. Such a unique surface cover configuration provides an exemplary experimental site with substantial theoretical and demonstrative value for investigating high-resolution remote sensing applications in land use mapping within extremely fragmented karst landscapes.

2.2. Datasets

This study utilizes VHR RS imageries (cloud-free data from dry seasons 2020–2022, spatial resolution ≤ 1 m) obtained through Google Earth Pro as the core data source for feature extraction. Through preprocessing procedures including orthorectification, radiometric normalization, and topographic shadow compensation, we effectively eliminated pixel displacement and radiometric distortions caused by dramatic karst terrain undulations. This process generated standardized imagery with both geometric precision and spectral fidelity, establishing a high-accuracy spectral-textural feature foundation for detailed land use classification.

A spatial constraint system was constructed through multi-source geodata integration: (1) road networks were extracted from OpenStreetMap to establish transportation corridors; (2) hydrological analysis using 30 m resolution SRTM DEM data delineated the Beipan River system framework, with canyon reach boundaries refined against high-resolution imagery to accurately characterize karst river incision features; (3) multiple topographic metrics, including relief amplitude, slope variability, and watershed boundaries, were derived from DEM data; and (4) national geographic census data were incorporated to obtain coarse-grained land use information. Through spatial overlay analysis, the study area was partitioned into geographic subregions, transforming complex mountainous terrain into computational units while introducing prior knowledge constraints.

3. Methods and Experiment

3.1. Methods

Influenced by geological structure, topography, and geomorphology, the distribution of land use objects in karst mountains is inherently fragmented, and spatial structure exhibits significant heterogeneity. This complexity introduces substantial uncertainty in remote sensing image interpretation, making it challenging to achieve fine-grained mapping solely through unified classification scales and algorithms. Consequently, this study proposes a geographical zoning and stratified object extraction framework. As illustrated in Figure 2, the approach involves three sequential stages. First, leveraging geographic zoning principles, we implement multi-scale spatial zoning from coarse to fine scales, establishing prior constraints for subsequent object extraction. Subsequently, considering the visual differences among various land use types, a stratified extraction scheme of objects is constructed. Then the model of the DL network is utilized to enable stratified object extraction. Finally, extracted objects are integrated and post-processed to generate a fine-grained land use map of karst mountains. Subsequently, spatial structure analysis is conducted based on landscape metrics.

3.1.1. Geographical Zoning

Zoning is a fundamental method in geological research to express the patterns of regional differentiation, reflecting a deep understanding of the spatial organization characteristics of the land-earth surface system. It aims to capture and quantify the inherent spatial dependence and heterogeneity characteristics of geographical phenomena, promote the merging of similar and adjacent units to form homogeneous regions, and divide dissimilar units to construct boundaries of heterogeneous phenomena. This process is essentially an important pathway for the transformation of surface systems from disorder to order. Overall, geographical zoning is not only increasing the sample diversity via a delineation of geographical phenomena but also reflecting a deep understanding of spatial relationships and regional characteristics.

In this study, geographical zone refers to the process or outcome of dividing or simplifying complex land surfaces based on the spatial distribution patterns and interaction mechanisms of objects [34,35]. Establishing multi-level geographical zones not only reveals regional differentiation patterns in complex mountainous systems but also minimizes errors in subsequent object extraction. By integrating zoning standards and visual cognition, the target area is stratified hierarchically into three levels: wide-area space, natural zone, and functional zone. The wide-area space represents our study area. Natural zoning initially reduces the spatial heterogeneity of complex karst mountains. It constrains the delineation of functional zoning and simultaneously contributes to the construction of the sample set, ensuring the generation of representative and comprehensive samples. The homogeneity of geographic characteristics within a functional zone imposes inherent constraints on the object extraction process, thereby mitigating error propagation. It also improves the efficiency of land use object extraction.

First, guided by the functional zoning framework, the study area is divided into construction, ecological, and rural zones (as shown in Figure 3a), with boundaries defined by administrative limits, primary river systems, and major road networks. Specifically, construction zones serve as the primary area for human habitation and infrastructure, dominated by constructed and transportation land. Ecological zones represent areas with ecological protection functions, capable of providing ecosystem services and products, predominantly consisting of forests, grasslands, rivers, and lakes. Rural zones serve as designated production spaces where agricultural activities occur, mainly comprising croplands and orchards. For the specific zoning criteria, this study takes into account the insignificant differences in topography and climate among different regions within the experimental area, as well as the significant differences in characteristics between vegetation and non-vegetation. The spatial clustering method is used to achieve geographical zoning using the average NDVI values and coefficient of variation over multiple years of growing season.

Following natural zoning, linear control elements for subsequent zoning are extracted through three interconnected networks: (1) road networks formed by connecting road elements, (2) hydrological networks constructed from river system components, and (3) topographic networks derived from DEM data analysis (including ridgelines and gully lines). These networks are superimposed with administrative boundaries to subdivide the natural zone into internally homogeneous subregions. Finally, the land use products and the land survey data are synthesized to delineate functional zones containing agricultural zones, forest-grassland zones, impervious surface zones, water zones, and other zones. In the process of zoning, this study takes into account the appropriate size of subregional units and the reasonable modification of regional boundaries in combination with roads; the final spatial division of the region is achieved. Please see Figure 3b.

3.1.2. Stratified Object Extraction

The biological visual cognitive system exhibits typical hierarchical features. Attention in visual cognition selectively focuses on local information related to the current task while suppressing interference and redundant information. Therefore, for the visual interpretation task of RS images, this study adopts a stratified strategy to extract ground objects. Namely, treating one object as a positive sample and others as negative samples for class-by-class extraction, or sequentially treating one object as a positive sample and another object as a negative sample for pair-by-pair extraction.

Specifically, according to image cognition and target demand, the extracted target objects include two types of linear elements (roads and river systems) and six types of surface elements: building, water, cropland, orchard, forest-grassland, and other. Given the high spatiotemporal stability of roads and river systems, topological verification and manual correction on the acquired base geographic data are implemented to streamline processing. For the remaining six surface elements, we employ an object stratification strategy, where a single object is designed as a positive sample while other classes are negative samples for class-by-class extraction. In this step, the above geographical zoning is helpful to address the high heterogeneity characteristic of karst landscapes via data augmentation in each zone.

After systematically analyzing the morphology and spatial distribution patterns of target objects (see

Table 1. Description of stratified objects extraction for land use mapping.

Extraction Order	Land Use Objects	Image Characteristics	Representative Sample
1	Building	Regular shape, significant spatial clustering, large-scale independent monoliths with clear geometric boundaries in urban areas, and small-scale dense and continuous distribution in rural areas
2	Water	Homogeneous texture with blue-green hue, various shapes and sizes
3	Cropland	Clear boundaries, uniform texture, and limited adjacent class combinations
4	Orchard	Granular texture and defined boundaries
5	Forest-grassland	Complex vegetation cover characteristics, various shapes, and indistinct boundaries
6	Other	Bare land, wasteland, etc.

), a stratified extraction sequence of “building → water → cropland → orchard → forest-grassland → other” is established. The rationale is as follows. First, buildings typically demonstrate regular shape and significant spatial aggregation. Large-scale, independent monoliths with clear geometric boundaries in urban areas and small-scale, dense clusters or isolated monoliths with defined geometries in rural areas. Under the spatial constraints of the impervious surface zone, buildings receive the highest extraction priority. Second, despite variations in size and spatial morphology, water consistently displays homogeneous blue-green spectral signatures distinct from other features, warranting second priority. In the agricultural zone, cropland, with clear boundaries, uniform internal texture, and limited adjacent class combinations, is assigned the next priority. Orchard, distinguished by its granular texture and defined boundaries, receives higher priority. Forest-grassland is extracted last due to its complex vegetation cover characteristics and vague boundaries. The remaining land use types (e.g., bare land, wasteland) are categorized as “other”. This strategy follows the cognitive logic of progressing from distinct to ambiguous objects, prioritizing classes with high separability, and thereby optimizing feature interpretation efficiency in complex karst mountainous scenes.

After constructing the objects stratification extraction system, the hierarchical semantic boundary-guided network (HBGNet) [11] is employed to implement the extraction of all objects (as shown in Figure 4). This architecture jointly learns low-level boundary information and high-level semantic information, making it well-suited for the extraction task in this study. Compared to traditional semantic segmentation networks, it offers three key advantages. First, a Local-Global Contextual Aggregation (LGCA) module is developed to integrate local and global spatial information, enhancing feature object perception, particularly effective in complex terrain scenarios such as mountains and hills. Second, a Boundary Guided Module (BGM) is proposed to extract spatially explicit boundary features that guide and constrain multilevel semantic features, thereby reducing boundary fuzziness’ in extracted objects. Third, a Multi-Grained Feature Fusion Module (MGFM) is introduced to strengthen the representation of multi-granularity semantic information, thus enhancing the model’s ability to effectively extract small-size objects (e.g., broken cropland).

The loss function of the model consists of three parts: the mask prediction loss $l_{bce}$, the edge prediction loss $l_{nll}$, and the distance prediction loss $l_{mse}$.

$$l_{total}=l_{bce}+l_{nll}+l_{mse}$$

(1)

where $l_{bce}$ is measured by binary cross-entropy, $l_{nll}$ is calculated by negative log-likelihood loss, and the mean square error is used to calculate the $l_{mse}$:

$$l_{bce}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\left(y_m\log {\widehat{y}}_m+\left(1-y_m\right)\log \left(1-{\widehat{y}}_m\right)\right)},$$

(2)

$$l_{nll}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N\log }{p}_e\left(x;y_e(x)\right),$$

(3)

$$l_{mse}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N{\left({\widehat{y}}_d-y_d\right)}^2},$$

(4)

where i denotes the pixel index and N denotes the total number of pixels in the image. $y_m$ and ${\widehat{y}}_m$ represent the ground truth value and prediction result of object mask, respectively. $p_e\left(x;y_e(x)\right)$ denotes the predicted probability of the edge labeled value $y_e(x)$. $y_d$ and ${\widehat{y}}_d$ represent the label value and prediction result of distance.

3.1.3. Post-Processing and Analysis

To further improve the object extraction results, we designed a post-processing workflow. First, the image morphology processing techniques are further utilized to eliminate isolated pixels and small patches and fill small voids. Second, vectorization and topology checking are performed. All object extraction results are vectorized, and the Douglas-Puke algorithm is used to remove redundant nodes [29,36,37]. Topology errors are also checked manually to ensure the accuracy and consistency of the data. Finally, different object layers are merged. To avoid conflicts between different object layers, sequential masking and erasure operations are implemented during merging. Following post-processing, we achieved the fine-grained land use mapping. Based on the refined mapping results, we calculated landscape metrics of various land use objects-patch density (PD), mean patch size (MPS), and landscape shape index (LSI) to reveal the spatial heterogeneity of complex karst mountain areas.

3.2. Experiment

The ground truth dataset for land use mapping was constructed through manual visual interpretation based on high-resolution imagery and field survey data. After cross-validation by multiple interpreters, the positional accuracy of samples reached 95.2%. All samples were standardized to 500 × 500 pixels to comprehensively capture the intra-class spectral-textural variability of typical karst landforms (e.g., bare rocks, crevice vegetation). To enhance model robustness, data augmentation strategies including random horizontal/vertical flipping, ±30° rotation, and HSV color space perturbations (hue ± 0.1, saturation ± 0.2, value ± 0.1) were implemented, expanding the original sample size by a factor of 5. This process yielded a final training-validation set containing 50 × 5 samples. In addition, the experiment was implemented using the PyTorch 1.7 framework, running on hardware configurations comprising an NVIDIA GeForce RTX 3080 GPU (10 GB VRAM) and an Intel i7-9300 CPU, accelerated by CUDA 12.0 libraries. Training parameters were configured with a batch size of 32 over 150 epochs, employing a cosine annealing learning rate scheduler (minimum learning rate 1 × 10⁻⁶).

4. Results and Analysis

4.1. Accuracy Evaluation Analysis

In this paper, to quantitatively assess the extraction accuracy of different object types, 8 sample areas, each measuring 1000 × 1000 pixels, were manually annotated across the entire study area. The evaluation was conducted using several standard metrics, including overall accuracy (OA), Kappa coefficient, mean intersection over union (mIoU), and F1 score. The calculation formulas are as follows.

$$\begin{array}{c}OA={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}},\end{array}$$

(5)

$$\begin{array}{c}Kappa={\displaystyle \frac{{P_o-P}_e}{1-P_e}},\end{array}$$

(6)

$$\begin{array}{c}F1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}},\end{array}$$

(7)

$$\begin{array}{c}Precision={\displaystyle \frac{TP}{TP+FP}},\end{array}$$

(8)

$$\begin{array}{c}Recall={\displaystyle \frac{TP}{TP+FN}},\end{array}$$

(9)

$$\begin{array}{c}\mathit{mIoU}={\displaystyle \frac{TP}{TP+FP+FN}},\end{array}$$

(10)

where TP is true positives, TN is true negatives, FP is false positives, and FN is false negatives, $P_o$ is the observed accuracy, $P_e$ is the expected accuracy by chance. Furthermore, the standard deviation (SD) of mIoU as an evaluation metric to assess the stability and reliability of the proposed method’s performance.

The extraction accuracy of different land use types exhibits significant variation (see

Table 2. Accuracy of the extracted land use objects.

Land Use Type	OA	Kappa	F1	mIoU ± SD
Cropland	0.918	0.831	0.901	0.819 ± 0.627
Orchard	0.981	0.917	0.927	0.864 ± 0.639
Water	0.992	0.929	0.933	0.874 ± 0.363
Building	0.985	0.758	0.765	0.620 ± 0.437
Forest-grassland	0.888	0.690	0.760	0.613 ± 0.531
Other	0.865	0.435	0.502	0.335 ± 0.836
Overall	0.815	0.750	0.798	0.688 ± 0.572

). Among them, water achieves the highest recognition performance, primarily due to its distinctive spectral reflectance characteristics and clearly defined spatial boundaries. In terms of orchard extraction, the regular row-wise arrangement of trees caused by human cultivation results in distinct point-like or grid-like texture patterns in remote sensing imagery, making them easily distinguishable from other land use types. Cropland generally features well-defined block- or strip-shaped boundaries and relatively uniform internal texture, which contributes to its high extraction accuracy. However, during the early crop growth stage or fallow periods, the spectral characteristics of cropland become similar to those of bare land and other land types, reducing the model’s discriminative ability. Additionally, transitional zones at the edges of some cropland, especially where they interface with orchards or forest-grassland, lead to unclear boundaries and increase the difficulty of learning edge features. In the case of building extraction, although OA is relatively high, there is a noticeable issue of omission, particularly for small or low-rise buildings in mountainous areas. These features are prone to misclassification as bare land or roads due to limitations in the spatial resolution of remote sensing imagery and the lack of distinctive spectral characteristics. The accuracy of forest-grassland extraction is comparatively low; in addition to frequent confusion with cropland and orchards, the internal vegetation types are diverse, and the spectral reflectance characteristics vary greatly within the same class, making consistent modeling more difficult.

4.2. Spatial Structure Analysis

Using the methodology described in this study, a total of 890,832 land use objects were extracted and categorized into several major classes: water, building, cropland, orchard, forest-grassland, and other (see Figure 5). Spatial pattern analysis of these land use types reveals that forest-grassland areas dominate the landscape, accounting for 54.55% of the total area. Their distribution exhibits a distinct banded pattern, primarily concentrated in the mountainous regions of the south and north. These areas, characterized by high elevation (average elevation 1037 m) and steep slopes (average slope 24.39°), are unsuitable for agricultural cultivation and construction activities. As a result, extensive and contiguous forest-grassland ecosystems have developed, serving as critical ecological barriers that help maintain regional ecological security. Cropland, as the second most prominent land use type, covers 15.62% of the area and shows a marked spatial aggregation. It is mainly distributed in the flat terrain areas of the north and central regions (average slope < 13°). In the mountainous central and southern zones, influenced by geomorphic units such as river terraces and intermontane basins, cropland is scattered in small parcels. Orchards exhibit a clear edge-distribution pattern, accounting for only 3.54% of the total area. Their spatial characteristics include being predominantly located in the ecological transition zones between cropland and forest-grassland and being concentrated on gently sloping hilly terrain (slope 10–31.02°) with relatively low suitability for conventional agriculture. Compared to cropland, orchards show lower spatial continuity and have not formed large-scale contiguous areas. The road network density is significantly higher in the northern and central flatlands, forming a transportation system that connects settlements and croplands. This pattern is closely related to the region’s higher intensity of human activities. A building primarily composed of rural settlements demonstrates a dual-core spatial distribution pattern: within and along the edges of cropland matrices and along major transportation corridors. In the northern and central regions, due to population clustering and economic development needs, building is not only more densely distributed but also consists of larger individual units. As for water bodies, the main hydrological feature in the study area is a tributary of the Beipan River located at the boundary of the two counties. Additionally, the central and northern regions contain several scattered small water bodies (average area 0.00447 km²), but no large-scale continuous water systems with significant ecological functions are present.

From the perspectives of area, shape, and spatial distribution, four landscape metrics were selected to analyze the land use types, namely, PD, MPS, and LSI. The distribution and landscape metric values for each land class are shown in

Table 3. Landscape pattern indices for multi-patch land use types.

Land Use Type	Area (km2)	PD	MPS (km2)	LSI
Cropland	465.2	1580.29	0.00063	1.27
Orchard	105.6	188.15	0.00531	1.33
Water	4.3	222.09	0.00447	1.14
Building	30.3	4025.97	0.00025	1.09
Forest-grassland	1625.2	3.37	0.29646	4.32

. The analysis reveals that forest and grassland, serving as the core matrix and ecological barrier of the study area, are characterized by large patch areas and the lowest patch density (PD = 3.37), along with the highest mean patch size (MPS = 0.29646 km²). This extensive, continuous vegetation plays a vital role in maintaining regional ecological stability. Despite its high continuity, the forest-grassland class also shows a relatively high LSI (LSI = 4.32, second only to roads), indicating complex and irregular patch boundaries. These irregular edges suggest increased edge length, which may intensify edge effects. In contrast to the good connectivity of forest-grassland areas, all other land use types dominated by human activities—buildings, cropland, water bodies, orchards, and roads—exhibit varying degrees of fragmentation. Building shows the highest fragmentation, with the greatest patch density (PD = 4025.97) and the smallest mean patch size (MPS = 0.00025 km²). This may reflect the presence of scattered settlements, small industrial facilities, or linear building expansions along roads. Cropland also demonstrates a high degree of fragmentation (PD = 1580.29, MPS = 0.00063 km²), with a large number of small cropland patches, which may hinder large-scale agricultural production and mechanized operations, thereby increasing management difficulty and costs. The impact of karst terrain on the landscape index of cropland is mainly reflected in three aspects. Firstly, surface fragmentation leads to small and scattered patch areas in cropland, significantly reducing landscape connectivity. Secondly, the matrix heterogeneity formed by exposed rocks increases the edge density, exacerbating the ecological transition zone effect between cropland and non-agricultural land. Thirdly, the soil erosion caused by underground karst cave systems manifested as a complex pattern of interlocking distribution between cropland, forest-grassland, and bare rock. This special topography limits the continuity of cultivation and alters the microtopography hydrology, ultimately resulting in irregular high-value characteristics of the landscape shape. Water (PD = 222.09), orchards (PD = 188.15), and roads (PD = 152.35) also display relatively high patch densities, indicating that these land types exist as numerous, relatively small patches. Regarding the LSI, which reflects the degree of human influence on land use types, building (LSI = 1.09), water (LSI = 1.14), cropland (LSI = 1.27), and orchards (LSI = 1.33) all exhibit low values, suggesting that these patch shapes are relatively simple and regular, typically the result of human planning and construction. In contrast, forest-grassland areas (LSI = 4.32) and roads (LSI = 6.21) have higher LSI values and more complex shapes. The intricate shapes of forest-grassland patches are largely driven by natural factors and their interspersion with other land use types, whereas the exceptionally high LSI of roads results from their inherently linear and networked structure, enabling maximum spatial connectivity and coverage with minimal area, thereby creating the most complex boundaries.

5. Discussions

5.1. Comparative Analysis

In this section, we demonstrate the enhanced discriminative capability of the selected model in feature extraction and evaluate the accuracy improvement achieved through the hierarchical strategy over conventional multi-class approaches. Furthermore, we emphasize the integration of prior categorical constraints via land use zoning, which imposes domain-specific categorical and boundary restrictions on the mapping results, effectively mitigating counterintuitive spatial distributions and intermixing of various objects caused by model uncertainties.

First, as shown in Figure 6, qualitative evaluations revealed distinct visual superiority of the selected model over conventional architectures (i.e., Unet [38] and DeepLabV3+ [39]) for both single- and multi-type object recognition, particularly in terms of geometric fidelity, boundary completeness, and spatial consistency. In complex karst landscapes, the HBGNet effectively identifies and retains the delicate shape characteristics of different objects, producing clear and intact boundaries in the extraction results, while the outputs of conventional models may present significant boundary blurring and breakage due to their limited adaptability to heterogeneous textures and irregular shapes. In addition, by fully utilizing boundary information, the HBGNet enhances the internal filling of objects, avoiding excessive holes and fragmented areas, thereby making the extraction results more visually coherent and complete.

The stratified extraction framework further enhanced visual plausibility by reducing inter-class confusion through a dedicated extraction order, enabling task-specific models to focus on discriminative spatial-textural cues. Unlike a single multi-class model, which often conflates spectrally similar objects, the stratified approach prioritized high-confidence objects (e.g., buildings) before progressively resolving illegible ones. This sequential refinement ensured that dominant features were fully captured without interference from adjacent objects, resulting in logically ordered and topologically consistent object hierarchies. The improvement in detection accuracy is particularly pronounced for orchard objects, which may share similar spectral-textural signatures in VHR imagery with both forest-grassland and cropland areas. The single-class model for garden land successfully captures garden land parcels as discrete, geometrically regular units with aligned planting rows, while accurately excluding adjacent forest-grassland objects with chaotic canopy structures and cropland objects with relatively homogeneous textural characteristics. As demonstrated in

Table 4. Comparative accuracy comparison of unified and stratified frameworks.

Mode	OA	Kappa	F1	mIoU
Unified extraction by a single model	0.787	0.675	0.745	0.603
Stratified extraction by independent models	0.815	0.750	0.798	0.688

, the stratified strategy—deploying specialized models for individual land cover categories followed by spatial aggregation—achieves statistically significant improvements across all accuracy metrics compared to a monolithic multi-class model. The marked enhancement in mIoU scores underscores optimized spatial morphology in extraction results, attributable to reduced class boundary ambiguities and intra-category spectral variance.

Second, the geographical zoning strategy, supported by land use prior knowledge, effectively mitigated extraction errors and topological inconsistencies during object vectorization, as shown in Figure 7. Through the incorporation of land use patterns as spatial constraints, the method suppressed errors where extracted objects conflict with known land use patterns. Additionally, during the raster-to-vector conversion, the integration of prior knowledge resolved ambiguities where linear features (e.g., roads) or fragmented woodlands were erroneously enclosed by adjacent cropland polygons. The refinement mechanism operates specifically on land cover extraction outputs, achieving a collective elimination rate of 7.2% of misclassified objects. This systematic false-positive reduction demonstrates the algorithm’s capability to rectify boundary inaccuracies and spectral confusion artifacts inherent in initial segmentation outputs. However, it must be pointed out truthfully that, as the geographical zoning depends heavily on expert knowledge, it is necessary to address the potential subjectivity introduced by this process. This issue is the implications for reproducibility and generalizability. As the reliance on expert knowledge for zoning and extraction orders introduces subjectivity, it is a potential improvement to design semi-automatic or automated or semi-automated zoning techniques based on multi-source data and quantitative criteria.

5.2. The Value of Mapping Results

By providing high-resolution, high-precision land use maps, micro-topographic unit characteristics, and information on their dynamic changes, a breakthrough improvement in the quality of fundamental ground data can be achieved. This not only offers essential baseline data for studying the highly heterogeneous ecological functional spaces in karst regions but also enhances the spatial representation capabilities of the ecological value accounting system (including functions such as water conservation, soil retention, and carbon sequestration).

This work improves the identification accuracy of unique karst geomorphological units and optimizes the foundation for ecological value assessment. Through the digital representation of the “three-dimensional” structure of karst areas, precise delineation of micro-topographic units such as peak-cluster depressions, dissolution fractures, and sinkholes can be achieved. This enables ecological value assessment models to perform calculations based on actual surface conditions. Meanwhile, the integration of geographical and natural zoning as a technical approach can address the issue of ecological value assessment suitability caused by the high heterogeneity of karst regions. Differentiated parameters can be applied to different zonal units, thus avoiding systemic errors caused by parameter generalization in traditional methods. For instance, in the empirical study of the Huajiang section of the Beipan River Basin in Guanling County and Zhenfeng County, the use of geographical zoning techniques allows the incorporation of prior knowledge such as carbonate rock purity, degree of fracture development, and dissolution intensity into the mapping process. This makes it possible to accurately quantify the differences in ecological function of the same vegetation type under different karst conditions, thereby making the results of ecological value assessment more consistent with the regional characteristics of karst mountain areas.

This work also enhances object recognition capability and spatial detail preservation to support precise ecological governance decisions. In this study, the land surface is classified into different land use objects, and the spatial relationships and hierarchical structures among land use features are established to simulate the complex ecological processes of karst regions. This enables the identification of fine-scale ecological functional units that are difficult to distinguish using traditional methods, such as remnant vegetation patches within karst rocky desertification control areas, early-stage vegetation recovery zones, and small-scale land use units. In subsequent ecological value assessments, calculations can be made at the level of specific object units, allowing results to reflect the unique spatial differentiation patterns of karst regions. This facilitates effective distinction of functional differences within similar land use types and is more suitable for micro-level ecological value assessment and analysis at the county or watershed scale. As a result, ecological value assessments can be accurate down to specific management units, providing direct support for scientific decision-making and precise policy implementation in regional ecological and environmental governance. In fact, this study uses static VHR RS imagery from 2020 to 2022 to generate object-based land use maps. Incorporating multi-temporal analysis could enhance understanding of land use changes and ecological dynamics. This can comprehensively and meticulously grasp the dynamic changes in natural resources and ecological environment and guide the formulation of effective management strategies.

6. Conclusions

This study presents an advanced framework that synergistically integrates VHR RS imageries and prior knowledge for land use mapping in complex karst mountain areas. The proposed methodology innovatively combines the geographical zoning and stratified object extraction with the deep learning networks. By partitioning the study area into computationally confined subregions and hierarchically extracting distinct geographic objects (e.g., buildings, water, croplands, orchards, and forest-grassland), the framework significantly enhanced object discrimination and boundary delineation, exhibiting significantly enhanced performance. The experimental implementation in the Beipanjiang River Basin (Guanling and Zhenfeng counties) demonstrated robust performance, achieving an OA of 0.815, a Kappa coefficient of 0.750, an F1 of 0.798, and an mIoU of 0.688 overall. Notably, the heterogeneity of the extraction results for different objects is indicative of precision, which is significantly related to their image features. Through this approach, we successfully identified and mapped 890,832 objects, with forest-grassland areas predominantly dominating the landscape, followed by cropland. The spatial distribution patterns exhibited obvious variations among different land use types, and forest-grassland showed continuous distribution over large areas, while cropland demonstrated distinct scattered-clustered aggregation patterns, reflecting the influence of both natural factors (e.g., altitude) and human factors (e.g., anthropogenic accessibility). These findings not only establish a comprehensive database for the study region but also offer crucial a priori information for subsequent research initiatives. Furthermore, this methodology presents a valuable template for fine-grained land use mapping in other karst landscapes characterized by similar ecological fragility.

However, several limitations warrant consideration. The study is limited to a specific region in Guizhou Province. The methodology’s reliance on expert knowledge for determining geographical zoning criteria and stratified extraction order may introduce inherent uncertainties in the extraction process. While the stratified object extraction approach improved OA, the model’s generalizability to regions with different lithological or land use characteristics requires further validation. Given the universality of spatiotemporal heterogeneity in karst mountainous areas, scaling is also a crucial consideration for validating method generalization in other karst mountainous areas with different geomorphological and land use characteristics [40,41]. Hence, for the potential generalizability in extracting geographical zones and stratified objects via considering regional specificity, our next stage is to verify if the proposed method is applicable to other significant karst regions, such as other areas of Southwest China or the Mediterranean or Caribbean. In addition, while HBGNet performs well, its complexity in fully acquiring multiscale information of grasped objects from hybrid networks may hinder scalability in heterogeneous karst mountainous areas. Computational efficiency and potential for lightweight alternatives would be beneficial. Therefore, future work will focus on model adaptability through lightweight architectures, expanding validation to heterogeneous regions, and the integration of multi-temporal or multi-sensor data to enhance dynamic monitoring capabilities.