GIS-Integrated Semi-Supervised U-Net for Automated Spatiotemporal Detection and Visualization of Land Encroachment in Protected Areas Using Remote Sensing Imagery

1. Introduction

In recent years, threats to global ecotypes have become more and more severe, in particular, influenced by climate change and urban expansion, many of the society’s “restricted development” and “ecological redline” have been pillaged or destroyed. In the United States, extensive acreages of federal, state, and local woodlands, wetlands, and national parks are often threatened by invasive agriculture, road-building and development, and unlawful residential habitation. Such encroachment frequently happens in ecologically fragile margin zones, which can severely destruct the original habitat connections and hinder the migration course of wild animals and plants, thus significantly posing a threat to local ecological security, species diversity conservation, etc [1].

Specifically, in 2021, illegal encroachment into a number of the western US nature reserves elevated the danger of wildfire propagation, generating general public concern over our capacity to manage ecosystems [2]. Thus, the development of an automatic scheme to support remote sensing observation, spatial analysis and smart identification in order to realize the real-time and high-precision monitoring of illegal encroachments in protected areas becomes very appealing at this time.

As an Non-contact observation method, the remote sensing technique has played a significant role in monitoring the earth surface dynamic change. Remote sensing can obtain large-scale, time-series stable land surface image data by using multispectral or high-resolution sensors on board satellites or aerial platforms, which is extensively employed in land use classification, vegetation change detection, water monitoring and other applications [3]. When it comes to surveillance of protected areas, the advantage of remote sensing is that remote, manual or difficult-to-inspect areas are integrated and that the development of encroachment is available over various timespans.

There are also some open remote sensing image platforms in the United States such as Landsat, Sentinel, NAIP and so on supply covering multi-temporal observation data for decades, which have already been an important data data sources for the analysis of dynamic change and for the identification of ecological disturbance in protected areas [4]. In this context, remote sensing offers a cost-effective and high-frequency means to map micro-scale and incremental encroachment processes.

Geographic information system (GIS) provides a unified processing, analysis and expression platform for multi-source spatial data, which can integrate and manage heterogeneous information, such as remote sensing image, geographical boundary, geomorphology, policy partition. In the intrusion monitoring of protection area, GIS system is allowed not only to map the pixel-level intrusion mask generated by the model on space and finish the pre-design under the coordinate of geographic, but also, supported by GIS, to various more complicated space-time operations such as buffer analysis, neighborhood statistics and change trend analysis [5].

By superimposing and measuring the government’s protection red lines or historical management red lines, GIS can accurately identify spatial attribution and management degree for encroachment as well as the management authorities in charge [6]. Meanwhile, GIS visualization tools can also be applied to the evident demonstration of public opinion, ecological compensation evolution and policy optimization recommendation, which is a powerful means to support ecological intelligent governance.

2. Related Work

Gunawansa et al. [7] employed Sentinel-2 data and GIS technology in a protected region in Sri Lanka to reliably identify human-elephant conflict hotspots, primarily in the margins between forest and farmland through the NDVI time series analysis, kernel density estimation (KDE) of conflict incidents and the multi-factor overlay. Sharma et al. [8] analyzed the use cases of remote sensing and GIS tools or platforms (e.g., including GEE, ENVI, ERDAS, QGIS, etc.) in the management of natural resources and highlighted the importance of ground-based in-situ data on the validation and performance improvement of remote sensing/GIS model.

Mangewa et al. [9] used ground survey high-resolution images as training samples in Tanzania reserves, and used random forest and other algorithms to classify land cover on Sentinel-2 images, and obtained an Overall Accuracy of 94% and Kappa 0.92, which was better than the traditional Sentinel-2 classification. Wang et al. [10] constructed a combined GIS + UAV approach, first using GIS network analysis to evaluate cotton transportation pathways and environmental factors, and identified high-risk areas for volunteer cotton growth in 31 counties in southern Texas.

Lv et al. [11] constructed an efficient multi-source remote sensing forest swamp classification method for the Changbai Mountain Ecological Function Reserve, and designed a two-stage classification framework by integrating Sentinel-1/2, ALOS-2/PALSAR, Landsat-8, and environmental covariates such as topography and land surface temperature. Firstly, 94% of the artificial surface disturbances were eliminated by using the NDBI threshold, and then the optimized random forest classifierwas used to combine 42 features and generate a forest swamp map at 10 m resolution through ten-fold cross-validation. Nguyen et al. [12] used natural zoning information and remote sensing-GIS data to identify land use dynamics, and then realized short-term prediction through CA-Markov, and captured the nonlinear relationship between environmental and socioeconomic factors through the MOLUSCE module, which significantly improved the accuracy of future forest cover change simulation.

3. Methodologies

3.1. Semi-supervised pseudo-label optimization mechanism

In order to make full use of the large number of unlabeled samples in remote sensing images, we propose a semi-supervised training strategy based on pseudo-label generation and consistency constraints. The mechanism consists of three key steps: pseudo-label generation, consensus regularization, and finally joint optimization of the objective function.

Above all, the current model is used to predict the unlabeled sample $X_u$ to obtain the prediction probability graph ${\widehat{Y}}_u$, and then the pseudo-label is generated based on confidence threshold, as Equation 1:

$${\tilde{Y}}_u^{(i,j)}=\left\{\begin{array}{c}1,{\widehat{Y}}_u^{(i,j)}>\tau \\ 0,{\widehat{Y}}_u^{(i,j)}<1-\tau \\ ignore,otherwise\end{array},\right.$$

(1)

where ${\widehat{Y}}_u^{(i,j)}$ represents the prediction probability of the model at pixels $(i,j)$, $\tau$ is the confidence threshold (e.g., 0.9), and the region below this threshold will be treated as an uncertainty region and ignored.

This pseudo-labeling strategy ensures that only high-confidence regions are used for semi-supervised training, avoiding pseudo-label propagation noise. The pseudo-label region is trained using the standard cross-entropy loss function ${\mathcal{L}}_{pseudo}$, as shown in Equation 2:

$${\mathcal{L}}_{pseudo}=-\sum _{\left(i,j\right)\in {\Omega }_p}[{\tilde{y}}_{i,j}log{\widehat{y}}_{i,j}+(1-{\tilde{y}}_{i,j}\left)\mathrm{l}\mathrm{o}\mathrm{g}\right(1-{\widehat{y}}_{i,j}\left)\right],$$

(2)

where ${\Omega }_p$ represents the trusted region of the pseudo-label, ${\tilde{y}}_{i,j}$ is the pseudo-label value, and ${\widehat{y}}_{i,j}$ is the corresponding predicted value.

In order to further enhance the robustness of the model, a consistency loss term is introduced to constrain the prediction results of the model to be consistent before and after image perturbation, as shown in Equation 3:

$${\mathcal{L}}_{consistency}={\mathbb{E}}_{X_u}\left[{\parallel {\widehat{Y}}_u-{\widehat{Y}}_u^{\mathcal{T}}\parallel }_2^2\right].$$

(3)

where$\mathcal{T}(·)$ Represents image perturbation operations, such as random flipping, rotation, or color perturbations. Structured similarity masks are further introduced, emphasizing consistency in the edge region, as shown in Equation 4:

$${\mathcal{L}}_{sim}=\sum _{i,j}{M}_{i,j}·{\parallel {\widehat{y}}_{i,j}-{\widehat{y}}_{i,j}^{\mathcal{T}}\parallel }_2^2,$$

(4)

where $M_{i,j}$ represents the weight of the edge structure (e.g., generated by the Sobel operator), which can enhance the consistent learning of the model at the edge position. Overall loss function integrates monitoring mechanisms, as Equation 5:

$$\mathcal{L}={\mathcal{L}}_{sup}+{\lambda }_1{\mathcal{L}}_{pseudo}+{\lambda }_2{\mathcal{L}}_{consistency}+{\lambda }_3{\mathcal{L}}_{sim},$$

(5)

where ${\mathcal{L}}_{sup}$ is the supervised loss of the labeled image, and ${\lambda }_1$, ${\lambda }_2$, ${\lambda }_3$ are the weighted hyperparameters of three regular terms.

3.2. Multi-scale attention fusion module

In order to improve the traditional U-Net recognition ability of complex ground object edges in remote sensing images, a multi-scale dilated convolution and spatial attention fusion module is introduced to enhance the segmentation ability from two aspects: feature diversity and spatial saliency. Dilated convolutions with different expansion rates are used to extract multi-scale context information, as Equation 6:

$$F_{multi}=\sum _{k=1}^K{w}_k·Con{v}_{r_k}\left(F\right),$$

(6)

where $Con{v}_{r_k}$ denotes the dilated convolution with the expansion rate $r_k$, $F$ is the input feature graph, $K$ is the number of scales, and $w_k$ is the learnable weight. Combined with global mean pooling and maximum pooling, the guided model focuses on the salient region, which is expressed as Equations 7 and 8:

$$A_{spatial}=\sigma \left(Con{v}_{1\times 1}\left(AvgPool\left(F\right)+MaxPool\left(F\right)\right)\right),$$

(7)

$$\tilde{F}=F·A_{spatial},$$

(8)

where $\sigma$ represents sigmoid activation, and $·$ represents element-by-element multiplication. This mechanism effectively enhances the response of the model to edge and weak features, and improves the segmentation accuracy.

The mask map of the final output of the model needs to be registered to the GIS spatial coordinate system for dynamic change analysis, as Equation 9:

$$\left(x_g,y_g\right)={\mathcal{T}}_{geo}\left(i,j\right)=\left(x_0+j·r_x,y_0-i·r_y\right),$$

(9)

where $(x_0,y_0)$ is the geographic coordinate of the upper left corner of the image, $(r_x,r_y)$ is the image resolution (e.g., 30 meters per pixel), and $(i,j)$ is the pixel index.

The Figure 1 shows the overall architecture of the semi-supervised enhanced U-Net (SS-EU-Net) model proposed in this paper, which mainly includes input remote sensing images, encoder module, self-supervised pre-training branch, multi-scale attention fusion module and decoder output. The input on the left side of model is a remote sensing image with geographic coordinates, and after multi-layer coding to extract features, enters module for scale fusion and saliency enhancement.

At the same time, the self-supervised branch at the top uses a convolutional neural network to model the structure information of the unlabeled image, and is used to generate pseudo-labels and consistency regularization loss, so as to enhance the stability of training. Finally, the decoder generates a pixel-level mask of the encroachment area, and improves the utilization rate of unlabeled data through the pseudo-label mechanism, so as to realize the automated.

4. Experiments

4.1. Experimental setup

In this section, we utilize Landsat 8 from USGS as the primary dataset to address the commonly ecological reserves in the United States. With a spatial resolution of 30m and abundant multispectral bands and multi-year time series information, the dataset can be used to monitor land cover change, illegal invasion, etc. According to the characteristics of different bands in remote sensing images, we conduct a systematic evaluation on the accuracy and robustness of the model for protected area encroachment detection by calculating the pixel-level masks of model output and the real labels visually. We selected four representative methods in the field of remote sensing semantic segmentation and semi-supervised learning as benchmarks:

• U-Net is a classical fully convolutional neural network structure, which is widely used in semantic segmentation tasks. It adopts an “encoder-decoder” symmetrical structure, and uses skip-connection to fuse low-level details with high-level semantic information.
• Mean Teacher is a mainstream semi-supervised learning framework, which constructs two structures: “student model” and “teacher model”, in which the parameters of the teacher model are the exponential moving average of the student model at different training stages.
• DeepLabV3+ is an advanced semantic segmentation model that introduces Empty Space Pyramid Pooling (ASPP) and encoder-decoder architectures, and has strong multi-scale context modeling capabilities.
• FixMatch combines a pseudo-labeling mechanism with a strong and weak data augmentation strategy. It uses weak enhancement to generate pseudo-labels, strongly enhances input images, and filters labels with confidence thresholds to guide network training.

4.2. Experimental analysis

Intersection over Union is used to measure the degree of overlap between the prediction mask and the true mask.

The results of Figure 2 show that with the increase of the proportion of labeled data, the IoU of each method increases, SS-EU-Net always maintains the highest level, especially when the annotation ratio is less than 40%, and its IoU is about 3–5 percentage points higher than that of the second place FixMatch, showing the significant advantage of semi-supervised pre-training and multi-scale attention fusion in the few-shot scenario. At high annotation ratios (≥70%), the performance of all methods tends to be saturated, but SS-EU-Net still leads by about 2–3 percentage points, demonstrating that it can not only make effective use of limited labels, but also further improve segmentation accuracy under large-scale supervision.

Figure 2. IoU Comparison Across Methods With Labeled Data Ratio.

Figure 2. IoU Comparison Across Methods With Labeled Data Ratio.

Figure 3. ‘F1 Score Comparison Across Methods With Training Epochs.

Figure 3. ‘F1 Score Comparison Across Methods With Training Epochs.

As can be seen in Figure 2, SS-EU-Net always maintained the highest level and converged the fastest throughout the training process. In the 30–50 rounds, the F1 value is already above 0.80, while other methods require a higher number of rounds to reach the same level. FixMatch came in second and also showed a strong semi-supervised advantage in the mid-to-low rounds; DeepLabV3+ and Mean Teacher gradually approached SS-EU-Net after a high number of rounds (≥70 rounds), but always lagged behind by about 0.02–0.03, while U-Net converged the slowest and had the lowest overall F1 value.

As the cloud coverage gradually increases from 0% to 90%, Table 1 shows a clear downward trend in MCC for all methods, but SS-EU-Net has always maintained the highest level and the smallest decrease in all cloud occlusion conditions: SS-EU-Net reaches 0.88 when there is no cloud, while U-Net only reaches 0.82, and when cloud coverage increases to 90%, SS-EU-Net remains around 0.44, while U-Net has dropped to 0.35.

5. Conclusion

In conclusion, this paper proposes a semi-supervised enhanced U-Net (SS-EU-Net) model, which integrates self-supervised pre-training, pseudo-label optimization and multi-scale attention fusion modules to achieve efficient and accurate pixel-level intrusion area segmentation, and completes GIS layer overlay and spatiotemporal evolution monitoring through geographic coordinate registration. Large-scale experimental comparisons show that SS-EU-Net is significantly better than U-Net, Mean Teacher, DeepLabV3+ and FixMatch in IoU, F1, MCC and other indicators, and shows excellent robustness in the context of high cloud cover and low labeling. In the future, the work will be expanded to multi-phase and multi-source data fusion, and online incremental learning and lightweight deployment will be explored to further improve the model generalization capability and real-time monitoring efficiency, and help intelligent ecological protection and decision support.