Assessment of Urban Rooftop Photovoltaic Potential Based on Deep Learning: A Case Study of the Central Urban Area of Wuhan

Abstract

Accurate assessment of urban rooftop solar photovoltaic (PV) potential is critical for the low-carbon energy transition. This study presents a deep learning-based approach using high-resolution (0.5 m) aerial imagery to automatically identify building rooftops in the central urban area of Wuhan, China (covering seven districts), and to estimate their PV installation potential. Two state-of-the-art semantic segmentation models (DeepLabv3+ and U-Net) were trained and evaluated on a local rooftop dataset; U-Net with a ResNet50 backbone achieved the best performance with an overall segmentation accuracy of ~94%. Using this optimized model, we extracted approximately 130 km² of suitable rooftop area, which could support an estimated 18.18 GW of PV capacity. These results demonstrate the effectiveness of deep learning for city-scale rooftop mapping and provide a data-driven basis for strategic planning of distributed PV installations to support carbon neutrality goals. The proposed framework can be generalized to facilitate large-scale solar energy assessments in other cities.

1. Introduction

The development of human society has always been deeply intertwined with energy utilization. For a long time, fossil fuels, as the dominant energy source, have supported global industrialization. However, their non-renewable nature has led to an increasing risk of resource depletion, accompanied by air pollution, ecological damage, and extreme climate events, posing severe challenges to human living environments. Accelerating the energy transition and building a sustainable, low-carbon energy system have become a global consensus.

The large-scale development of renewable energy is the core pathway for energy transition [1,2,3,4]. In recent years, renewable energy has accounted for approximately 60% of global new electricity generation, with over 130 countries and regions, including the European Union, the United States, and Japan, proposing carbon neutrality goals. On 22 September 2020, at the 75th United Nations General Assembly, China announced its “dual carbon” goals: “striving to peak carbon dioxide emissions by 2030 and achieve carbon neutrality by 2060”. This systematic strategic deployment demonstrates China’s determination to drive the energy transition. Solar energy, as a core component of renewable energy, is characterized by its infinite reserves, cleanliness, and strong regional adaptability, making it a key driver of the global energy transition.

Photovoltaic power generation, the primary form of solar energy utilization, includes centralized and distributed modes. Among these, distributed rooftop PV systems, integrated with urban buildings, do not require additional land resources and offer advantages such as proximity to the grid, reduced transmission losses, and flexible adaptation to urban spaces, showcasing unique potential. Internationally, the United States launched the “Million Solar Roofs Initiative” in 1997, Japan formulated the “Roof Solar Plan” (aiming for full coverage of applicable rooftops by 2030) [1], while China implemented the “Golden Sun Project” in 2009, promoting urban rooftop PV projects through fiscal subsidies. These policies share a common goal: to unlock the critical value of rooftop PV in optimizing energy structures through meticulous assessment and large-scale deployment.

Accurately assessing urban rooftop PV potential is crucial for China’s energy development. The coefficient method, as a mainstream approach, derives rooftop area coefficients from statistical data such as population and building footprint area. While cost-effective, it suffers from significant errors. For example, Gutschner et al. [5] estimated PV-suitable rooftop areas based on building footprints but overlooked spatial features such as building height and roof tilt. Peng J et al. [6] determined through field measurements that the ratio of rooftop area suitable for PV installation to ground area was approximately 0.45, but this coefficient fails to reflect differences in urban functional zoning. Wiginton et al. [7] sampled 10 out of 109 census divisions and established a relationship between rooftop surface area and population (70.0 m² per capita) but did not account for the impact of high-rise building proportions on actual usable space. Such methods face challenges in megacities like Wuhan—where central urban areas feature high building density and a predominance of high-rise residential buildings—as single coefficients struggle to capture complex urban textures. Additionally, geographic information system (GIS)-based methods are effective for studying urban solar potential. Li Ko et al. [8] estimated Taiwan’s potential by combining GIS with data on arable land, grassland, and built-up areas. Geeta Bhatta et al. [9] were the first to assess Nepal’s potential for ground-mounted, rooftop, and agrivoltaic PV systems, finding significant potential for agrivoltaic applications on farmland. However, GIS-based approaches often face data and computational challenges at large scales, whereas deep learning algorithms can more efficiently handle such tasks. With advancements in machine learning algorithms, image recognition accuracy has reached a viable level. Teng Zhong et al. [10] employed DeepLab v3 to calculate rooftop PV potential in Nanjinge, but their work focused primarily on 2D image analysis without considering urban functional zoning differences. Guannan Li et al. [11] developed a “GIS-deep learning” fusion framework for semantic segmentation of rooftops, achieving improved accuracy after classifying roofs into commercial, industrial, and residential categories. Jian Xu et al. [12] created a spatiotemporal feature-encoded deep learning model integrating 3D building morphology, solar position, and meteorological data for Futian District, Shenzhen, enabling dynamic shadow prediction and computational efficiency breakthroughs. However, it relied on high-precision LiDAR point cloud data, incurring high costs, and did not refine functional zoning differences based on land-use planning. Lodhi et al. [13] developed a city-wide assessment framework using Mask R–CNN and DeepLabv3 to extract solar panels and suitable rooftops from satellite imagery. They delineated thousands of solar arrays and quantified Islamabad’s current and future PV generation potential. Lodhi et al. [14] further advanced urban solar potential mapping by integrating deep learning with atmospheric radiative transfer modeling, enabling precise calculation of PV yield under varying cloud fraction, optical depth, and aerosol conditions. Zech et al. [15] employed a deep active learning approach to reduce annotation effort while maintaining high segmentation accuracy. Their entropy-based strategy achieved robust rooftop PV detection with only 3% of the labeled data, yet their focus on data efficiency overlooked model adaptability across diverse urban landscapes. Ding et al. [16] proposed an Adversarial Shape Learning Network (ASLNet) that addresses issues of occlusion and boundary ambiguity in building extraction from very high-resolution remote sensing images through adversarial learning strategies and a CNN shape regularizer, thereby improving segmentation accuracy. He et al. [17] constructed the ST-UNet framework, which embeds the Swin transformer into UNet to form a dual-encoder structure, enhancing the performance of remote sensing image semantic segmentation through spatial interaction, feature compression, and relational aggregation modules. Wang et al. [18] demonstrated through five experiments that the emotion of awe can reduce people’s territoriality, thereby enhancing sharing intention, and this effect is more significant in contexts of high perceived similarity and promotion focus. Jung et al. [19] proposed a boundary enhancement semantic segmentation method that combines a holistically nested edge detection (HED) unit and a boundary enhancement (BE) module to improve the boundary accuracy of building extraction from remote sensing images. Zheng et al. [20] proposed the FarSeg++ foreground-aware relation network, which models the foreground from the perspectives of relation, optimization, and objectness to alleviate false alarms and foreground-background imbalance in geospatial object segmentation from high-spatial-resolution remote sensing images. However, none of these documents involve the assessment of photovoltaic potential.

Existing deep learning methods for rooftop PV potential assessment heavily depend on dataset characteristics and algorithm compatibility. Public datasets often suffer from insufficient resolution or lack regional architectural features (e.g., Asian arcades, high-rise corridors), limiting model generalizability. The primary objective of this study is to develop a high-resolution imagery-based method for assessing urban rooftop PV potential by leveraging semantic segmentation algorithms in deep learning. The key novelties of this work include: (1) constructing a high-resolution local rooftop dataset capturing unique urban forms; (2) systematically comparing multiple segmentation models (DeepLabv3+ and U-Net variants) to identify an optimal approach; and (3) integrating rooftop segmentation results with urban functional zoning to refine PV capacity estimates. Urban climate models play a crucial role in accurately assessing photovoltaic potential by accounting for dynamic microclimatic factors, which are not captured in our current image-based geographical potential estimation. ENVI-met (version 3.1), a 3D microclimate model, can be used to analyze airflow between buildings, exchange processes between the ground surface and building walls, etc., and is applicable for analyzing rooftop photovoltaic performance under different shading and thermal conditions [21]. MITRAS mainly simulates turbulent flow and energy exchange within urban obstacle layers, coupling atmospheric transport and surface processes, and can accurately calculate solar irradiance attenuation caused by building clusters and aerosol effects [22]. MUKLIMO_3 is a high-resolution model for assessing urban heat load; it integrates land use classification, can resolve temperature inversions and ventilation effects, and is useful for studying how nocturnal cooling affects photovoltaic efficiency [23]. PALM is a large-eddy simulation model capable of resolving building-scale turbulence and vegetation–atmosphere interactions. Its application in Vienna has shown that it can effectively capture intra-urban thermal variability, including shading effects of adjacent buildings on rooftops [24]. This study focuses solely on geographical potential, i.e., the analysis of installable area for photovoltaic panels. Future research may consider integrating these models to achieve a dynamic simulation of actual power generation.

2. Methodology and Model

2.1. Overall Research Framework

This study utilizes publicly available and freely accessible Esri World Image high-resolution imagery to preprocess the data and select sample data, constructing a dataset for roof segmentation in the central urban area of Wuhan. The research framework of the study is shown in Figure 1. First, the imagery of the central urban area of Wuhan with a resolution of 0.5 m/pixel was downloaded. The imagery was then cropped into 21,498 images of 512 × 512 pixels. A total of 1165 images were selected as sample imagery, for which building roof contours were manually annotated to construct the dataset. The DeepLab v3+ and U-net network architectures in deep learning were employed, each paired with two different backbone feature extraction networks. Based on the constructed dataset, four semantic segmentation models were trained: DeepLabv3+_Mobilenet V2, DeepLabv3+_Xception, U-net_VGG16, and U-net_Resnet50. Subsequently, a comparison of roof extraction performance and evaluation metrics was conducted. The model with the best performance was used to automatically extract the roof contours of seven districts in the central urban area of Wuhan. The area of the extracted roofs was calculated using the pixel value screening method, which was then converted into photovoltaic (PV) installed capacity and annual power generation to assess the building rooftop photovoltaic power generation potential in the central urban area of Wuhan. This study evaluates the geographical PV potential—i.e., the theoretical installable capacity based on available roof area and assumed efficiency—and does not model actual solar irradiance, temporal performance, or shading effects.

2.2. Study Area

This study was conducted in the central urban area of Wuhan City, Hubei Province, China. As the provincial capital and one of China’s most historically significant cities, Wuhan’s central urban district comprises seven administrative regions: Jianghan, Hongshan, Hanyang, Qingshan, Wuchang, Qiaokou, and Jiang’an Districts. More than half of the city’s population is concentrated in this central area, with recorded permanent residents reaching 6.975 million in 2022 across a total area of 955.15 km². Characterized by a subtropical monsoon humid climate, the region experiences cold winters and hot summers with consistently high humidity levels. Meteorological data for 2022 indicates an annual average temperature of 18 °C and relative humidity of 75%. Wuhan possesses abundant solar energy resources, with annual solar radiation ranging from 104 to 113 kcal/cm² and approximately 1800 sunshine hours annually.

With the continuous expansion of Wuhan’s economy and population, the demand for energy consumption has increased rapidly. In response, the city’s ongoing energy transition emphasizes enhanced energy efficiency and greater utilization of renewable sources, particularly solar energy. The city possesses favorable natural conditions for solar power generation. The central urban area of Wuhan has numerous residential building rooftops that can strongly support the construction of rooftop solar photovoltaic projects. Research on Wuhan is of great significance for the implementation and promotion of distributed rooftop solar PV projects in other urban areas of China. Figure 2 shows the 0.5 m/pixel resolution imagery of the central urban area of Wuhan City and its partial enlarged view, where rooftops can be distinguished with the naked eye.

2.3. Model Construction

Proper dataset partitioning is crucial during model development. From the total collection of 21,498 images covering the study area, we selected 1165 representative samples for annotation, which comprehensively encompass all seven administrative districts in Wuhan’s central urban area (

Table 1. Composition of the dataset.

Administrative District	Number of Segmented Images	Number of Images Selected as Samples	Geographic Location
Jianghan District	682	243	Along the Yangtze River
Qiaokou District	981	324	West bank of the Han River
Qingshan District	1108	100	North bank of the Yangtze River
Wuchang District	2028	100	East bank of the Yangtze River
Jiang’an District	1795	100	Northeast section of Yangtze’s north bank
Hanyang District	2515	96	West bank at Yangtze–Han confluence
Hongshan District	12,389	202	Southeast of Wuchang
Total	21,498	1165	/

). Jianghan and Qiaokou Districts, as the core built-up areas of Wuhan with high building density and diverse rooftop types, were prioritized for a higher sample ratio to cover major rooftop forms in the central urban area. Other districts were sampled with approximately ~100 images each to ensure representativeness, while Hongshan District, being the largest in area, was supplemented with ~200 images to enhance spatial coverage. The labeled dataset was then partitioned into training and validation sets at a 9:1 ratio. The training set is utilized for the model’s parameter learning, enabling it to capture the characteristic patterns of rooftops from a large amount of rooftop imagery data. The validation set is employed to conduct real-time performance evaluations of the model during the training process, adjust hyperparameters, prevent overfitting, and ensure that the model maintains good generalization ability across rooftop imagery from different regions.

In this study, semantic segmentation algorithms in deep learning were employed to automatically extract rooftops and assess the rooftop photovoltaic geographical potential of rooftop PV of the central urban area of Wuhan at a spatial scale. Currently, there are numerous models for image segmentation. Among them, the fully convolutional neural network is the model with the best performance. Through a series of processes such as convolutional operations, forward and backward propagation, loss entropy calculation, and gradient descent, this model automatically learns to capture the semantic information of images, outputting the original remote-sensing images as predicted images with classification labels.

Currently, two network models, DeepLabv3+ and U-net, stand out for their excellent semantic segmentation performance. In this study, the DeepLabv3+ and U-net models were selected and paired with different backbone feature extraction networks to train the locally constructed dataset, respectively. The cross-entropy loss values between the predicted image labels and the true image labels were calculated. The building rooftop extraction effects were compared and analyzed. The optimal network model and parameters were retained to extract rooftops from the imagery of the central urban area of Wuhan.

DeepLabv3+, proposed by Google in 2018 [25], represents the most advanced version in the DeepLab series for image semantic segmentation. This architecture employs atrous convolution and atrous spatial pyramid pooling (ASPP) modules to expand the receptive field while enhancing feature extraction capabilities, consequently improving both segmentation speed and accuracy. The model has demonstrated exceptional performance in computer vision applications. Two backbone networks were evaluated: (1) MobileNetV2, which incorporates inverted residual blocks with linear bottlenecks to achieve high accuracy with reduced parameters [26]; and (2) Xception, where depthwise separable convolutions replace standard convolution operations, effectively decoupling spatial and cross-channel correlations [27].

The U-Net architecture was originally developed for medical image segmentation [28] as a fully convolutional network. Through successive convolution and pooling operations, it generates multi-scale feature maps containing high-level semantic information. The network’s simple structure enables fast computation while maintaining satisfactory performance even with limited training data. We evaluated two backbone networks: (1) VGG16, which stacks consecutive 3 × 3 convolutional kernels with pooling layers to construct deep convolutional neural networks [29], widely adopted in various computer vision tasks including semantic segmentation and image classification; and (2) ResNet50, introduced by Microsoft Research in 2015 [30], which utilizes residual connections to address vanishing gradient problems and batch normalization to accelerate training, enabling the development of substantially deeper network architectures.

In summary, the DeepLabv3+ and U-Net models adopt different semantic segmentation strategies: DeepLabv3+ employs atrous convolutions and an atrous spatial pyramid pooling module to capture multi-scale contextual information, whereas U-Net uses a symmetric encoder–decoder with skip connections to preserve fine-grained spatial details. The backbone networks also vary in depth and capacity: MobileNetV2 is a lightweight architecture optimized for efficiency, Xception and ResNet50 are deeper networks that can extract more complex features (with ResNet50’s residual design easing the training of very deep models), and VGG16 provides a deep but comparatively less efficient feature extractor.

2.4. Model Training

The annotated dataset, consisting of 1165 rooftop contour images, was simultaneously input into four models for training: DeepLabv3+_MobileNetV2, DeepLabv3+_Xception, U-Net_VGG16, and U-Net_ResNet50. Each model utilized pre-trained weights from its corresponding backbone (MobileNetV2/Xception for DeepLabv3+; VGG16/ResNet50 for U-Net) to avoid excessive randomness in the initial parameters. The input images, 512 × 512 pixels with three RGB channels, were derived from 0.5 m-resolution remote sensing imagery, matching the dataset construction in Section 2.1. The training process employed a two-stage strategy to optimize computational efficiency on a single NVIDIA A100 GPU (40 GB memory). During the initial 50 epochs (frozen phase), only the segmentation head was fine-tuned with a batch size of 8 and an initial learning rate of 0.001, while the backbone parameters remained fixed. In the subsequent 50 epochs (unfrozen phase), full parameter optimization was performed with a reduced batch size of 4 and a learning rate of 5 × 10⁻⁵. This strategy reduced total training time to approximately 4 h for 100 epochs. The Adam optimizer was adopted for its adaptive learning rate, and the Softmax function transformed output features into pixel-wise probability distributions. The dataset was split into training (1049 images) and validation (116 images) sets at a 9:1 ratio, with no independent test set due to limited annotated data. Model performance was evaluated via the validation set, which supported the selection of U-Net_ResNet50 as the optimal model (Accuracy = 0.94,

Table 2. Comparison of evaluation metrics for four models.

Model	mIoU	mPA	Accuracy
DeepLabv3+_Mobilenet V2	78.89	87.67	92.58
DeepLabv3+_Xception	77.36	88.38	91.60
U-net_ VGG16	82.57	89.84	94.08
U-net_ Resnet50	83.44	90.56	94.39

). Implementation was conducted using Python 3.7 with the PyTorch framework (torch 1.10), OpenCV-Python 4.1 for image processing, and GDAL 3.4.2 for geospatial operations.

Model convergence was evaluated through loss function monitoring, where decreasing values indicated improved prediction performance. The cross-entropy loss function quantifies the discrepancy between predicted labels and ground-truth labels by measuring the entropy difference between the predicted probability distribution and the actual distribution. Rooted in information theory, it assesses classification accuracy: smaller cross-entropy values indicate closer alignment between predicted and real distributions, guiding the model to learn feature distinctions for precise pixel-level rooftop segmentation. As shown in Figure 3, all four models exhibited high initial loss values that gradually decreased and eventually stabilized during training. For the DeepLabv3+_Mobilenet V2 and DeepLabv3+_Xception models, the total loss values decreased rapidly in the early stage, showing a distinct fluctuating downward trend. After approximately the 40th epoch, the rate of decrease slowed down, but they continued to decline. Around the 60th epoch, both values approached relatively stable levels. Among the U-Net architectures, the VGG16 variant displayed significant early loss reduction until epoch 20, followed by more gradual improvement. Notably, the ResNet50-based model outperformed others, achieving the fastest initial loss reduction and reaching stability by epoch 35—the earliest convergence among all models. This architecture also maintained superior validation loss performance, demonstrating exceptional training efficiency and stability. The ResNet50 model’s robust performance, characterized by stable convergence with fewer epochs and consistently low validation loss, suggests superior feature learning capability compared to other configurations.

2.5. Model Comparison

Three images containing different types of rooftops were used to predict with the trained models, and the building rooftop extraction effects of different models were compared. Figure 4, Figure 5 and Figure 6 are the comparison charts of the rooftop extraction effects of the four models for low-rise, medium-height, and tall buildings, respectively. It can be seen that when the two DeepLabv3+ series models extracted building rooftops, there were many edge-blurred pixels, resulting in unclear extracted rooftop boundaries or even failure to detect boundaries and pixel adhesion. In the prediction map of the low-rise rooftop in Figure 4, the path beside the rooftop was obviously misclassified as a rooftop. Therefore, the segmentation performance of the two U-Net series models was significantly higher than that of the two DeepLabv3+ series models. Compared with the U-Net_VGG16 model, the U-Net_ResNet50 model has clearer segmentation boundaries, a cleaner overall image, and is the only model that can completely identify the rooftops in the second-to-last row at the lower-left corner of Figure 6, demonstrating stronger detail processing capabilities.

The mean intersection over union (mIoU): mIoU is a comprehensive metric for evaluating segmentation accuracy, defined as the average of Intersection over Union (IoU) values across all classes. For our binary task (rooftop vs. background), IoU measures the ratio of correctly identified rooftop pixels to the total pixels involved in model predictions and actual rooftop regions. mIoU is the average of IoU values for the rooftop and background classes. Mean pixel accuracy (mPA): mPA represents the average per-class pixel classification accuracy. It is calculated by separately computing the accuracy for rooftop pixels (correctly classified rooftop pixels divided by total actual rooftop pixels) and background pixels (correctly classified background pixels divided by total actual background pixels), then averaging these two values to assess the model’s comprehensive classification performance across classes. Overall accuracy: Overall accuracy is defined as the proportion of correctly classified pixels to the total number of pixels, computed as the sum of diagonal elements in the confusion matrix divided by the total pixel count. This metric reflects the model’s overall correctness in classifying all pixels. These evaluation metrics are constructed based on the confusion matrix, a commonly used method in machine learning for evaluating model performance.

Table 2 presents a comparison of the evaluation metrics for the four models. The U-Net_ResNet50 model achieved superior scores across all standard segmentation metrics: mean intersection over union (mIoU) of 0.83, mean pixel accuracy (mPA) of 0.91, and overall accuracy of 0.94. These results compare favorably with existing approaches in the literature, including Teng et al.’s [10] spatially-optimized DeepLabv3 (0.92 accuracy) and Chen et al.’s [31] Wuhan-specific model (0.81 accuracy). The demonstrated accuracy of 0.94, combined with its robust performance on challenging test cases, confirms U-Net_ResNet50 as our optimal choice for comprehensive rooftop mapping across Wuhan’s central urban area. This selection reflects both quantitative superiority and qualitatively better handling of complex urban roof geometries and spatial arrangements.

3. Results and Discussion

3.1. Roof Extraction Result

The trained U-Net_ResNet50 model was systematically applied to process all segmented imagery of Wuhan’s central urban area, with the extraction results organized by administrative district boundaries. As illustrated in Figure 7, the output visualization employs a standardized color scheme: red regions denote successfully extracted rooftops, while black areas represent background. District boundaries (Jianghan, Qiaokou, Qingshan, Wuchang, Jiang’an, Hanyang, and Hongshan) are clearly demarcated with yellow lines for precise geographical reference. The spatial distribution analysis reveals distinct urban morphological patterns across districts. Regarding the roof distribution, the roofs in Jianghan District and Qiaokou District are relatively densely distributed. A large number of red markers are clustered within these districts, indicating a compact building distribution. In Qingshan District, the roof distribution is relatively evenly spaced, yet less dense compared to the former two districts. The roof distribution in Wuchang District and Jiang’an District exhibits both clustered and dispersed characteristics. In some areas, red markers are concentrated, while in others they are sparser. Due to their large areas, the roofs in Hanyang District and Hongshan District are more dispersed. Red markers are relatively sparse in the vast areas, and they are only concentrated in some local parts. Although the annotated training dataset (1165 images) is relatively modest given Wuhan’s size, it was designed to capture diverse rooftop types across all districts. The high model accuracy achieved (~94%) suggests this dataset was sufficient for our study area, though a larger dataset could further improve generalizability.

3.2. Land Use Classification of Rooftops

By screening the pixel values of the red areas, the number of rooftop pixels identified in each district can be calculated. Combining with the resolution of the downloaded remote-sensing imagery, the capacity of the central urban area of Wuhan can be calculated using Equations (1) and (2):

$${area}_i=N_i\times {\beta }^2$$

(1)

$$capacity=\sum _i^7({area}_i\times D\times {\mathrm{U}}_i)$$

(2)

where ${area}_i$ represents the area of each type of rooftop, $N_i$ denotes the number of pixels of each type of rooftop, and β signifies the resolution of the remote-sensing imagery. The resolution of the remote-sensing imagery downloaded in this study is 0.5 m/pixel. Capacity refers to the installable photovoltaic capacity of each district. $D$ denotes the power density of PV panels. Referencing prior studies, $D$ was set at 200 W/m² (within the typical range of 100–200 W/m²). ${\mathrm{U}}_i$ represents the utilization coefficient of each type of rooftop. Drawing on existing research, the utilization coefficients for different roof types were adopted: residential land (0.8), commercial land (0.63), industrial land (0.92), educational land (0.58), public facility land (0.6), medical land (0.52), and historical and cultural land (0) [32].

The proportion of installable photovoltaic (PV) systems varies by roof type. In this study, rooftop land use was classified based on the functional zoning plans of Wuhan’s districts published by the Wuhan Municipal Bureau of Natural Resources and Planning. The technical workflow was as follows: First, computer vision techniques were employed to extract land-use features from the zoning plans. BGR color thresholds (with a tolerance of ±3) were defined for seven land-use categories (residential, commercial, etc.), and the findContours algorithm in OpenCV was used to extract polygon contours of each land-use type. Second, geographic registration was achieved through JGWX world files and the GDAL library, converting pixel coordinates to Web Mercator projection (EPSG: 3857). Third, spatial indexing was established, and the point-in-polygon topological detection algorithm (cv2.pointPolygonTest) was used to determine the land-use type of the central point of each target building. Figure 8 is a schematic diagram of the planning for Qiaokou District, Wuhan, where different colors represent distinct land-use types [33]. The land-use types in all central urban zoning plans were statistically categorized into seven classes: residential land, commercial land, industrial land, educational land, public facility land, medical land, and historical and cultural land. For complex plots, the Douglas–Peucker algorithm was applied for polygon approximation, and noise regions smaller than 50 pixels were filtered to ensure geometric validity of the contours.

Figure 9 presents the rooftop classification results for each district in the central urban area of Wuhan, illustrating the distribution of rooftop areas across seven administrative districts (Jianghan, Qiaokou, Qingshan, Wuchang, Jiang’an, Hanyang, and Hongshan) by different land-use types. The overall distribution pattern indicates that residential land occupies a substantial proportion in most districts, reflecting the fundamental role of residential functions in urban spatial utilization. For instance, the rooftop areas of residential land in Jiang’an and Hongshan districts are significantly larger than those of other land-use types, suggesting that these two districts serve as major residential hubs with relatively high population densities and strong housing demands. Qingshan District demonstrates exceptional industrial rooftop coverage, aligning with its strategic position as a key industrial base in Wuhan and reflecting the profound influence of industrial zoning on urban land utilization. Conversely, Hongshan District stands out for its extensive educational land rooftop areas, reflecting the concentration of educational resources and institutions in this region, likely attributable to the presence of numerous universities and educational facilities. As shown in

Table 3. Available area of each rooftop of each district.

Administrative District	Residential/m2	Available Area/W	Commercial/m2	Available Area/W	Industrial/m2	Available Area/W	Educational/m2	Available Area/W	Public Facility/m2	Available Area/W	Medical/m2	Available Area/W	Historical and Cultural/m2	Available Area/W
Jianghan District	4,466,552.48	3,573,241.99	2,248,221.37	1,416,379.46	0.00	0.00	650,915.65	377,531.08	591,300.68	354,780.41	255,754.07	132,992.12	0.00	0.00
Qiaokou District	5,735,469.56	4,588,375.65	2,340,714.44	1,474,650.10	0.00	0.00	1,460,604.03	847,150.34	316,761.27	190,056.76	434,007.88	225,684.10	66,121.07	0.00
Qingshan District	475,808.24	380,646.59	1,433,860.51	903,332.12	5,226,068.20	4,807,982.74	660,775.56	383,249.83	2,202,162.68	1,321,297.61	0.00	0.00	548,017.56	0.00
Wuchang District	6,755,023.95	5,404,019.16	4,719,065.03	2,973,010.97	0.00	0.00	2,606,654.83	1,511,859.80	537,872.26	322,723.36	1841.18	957.42	391,961.74	0.00
Jiang’an District	9,387,981.12	7,510,384.90	2,987,289.42	1,881,992.34	0.00	0.00	1,339,108.46	776,682.91	600,089.87	360,053.92	140,037.48	72,819.49	223,138.89	0.00
Hanyang District	7,365,718.99	5,892,575.19	2,800,428.91	1,764,270.21	2,721,292.16	2,503,588.78	701,056.75	406,612.91	2,820,345.80	1,692,207.48	0.00	0.00	226,296.15	0.00
Hongshan District	25,261,038.27	20,208,830.61	3,158,001.68	1,989,541.06	2,740,926.68	2,521,652.54	18,260,835.29	10,591,284.47	2,547,997.62	1,528,798.57	0.00	0.00	2,881,679.73	0.00

, the educational land in Hongshan District has an available area of 10,591,284.47 m², corresponding to a photovoltaic installed capacity of 1,237,681.03 kW. Commercial land generally accounts for limited rooftop areas across all districts. This may be attributed to commercial buildings prioritizing efficient space utilization and architectural aesthetics over expansive rooftops. Similarly, public facilities, medical, and historical–cultural land uses exhibit relatively small rooftop areas, indicating their low urban footprint or architectural and functional characteristics that constrain rooftop expanse. These observations inform urban planning strategies to optimize public resource allocation and infrastructure development based on district-specific land use characteristics. For photovoltaic potential assessment, the quantitative breakdown of rooftop areas by type and district enables precise capacity estimation, offering a scientific foundation for strategic renewable energy deployment and maximizing clean energy utilization efficiency. The data-driven approach facilitates tailored solar energy policies that account for the intrinsic relationship between urban functional zoning and renewable energy infrastructure planning.

3.3. Photovoltaic Potential in the Central Urban Area of Wuhan

Figure 10 is a visual heat map of the photovoltaic (PV) installed capacity per unit area in each district of the central urban area of Wuhan.

Table 4. Photovoltaic installation capacity results of each district.

Administrative District	Roof Area/m2	Geographical Potential (PV Installation Capacity/kW)
Jianghan District	8,212,744.25	1,170,985.01
Qiaokou District	10,353,678.25	1,465,183.39
Qingshan District	10,546,692.75	1,559,301.78
Wuchang District	15,012,419.00	2,042,514.14
Jiang’an District	14,677,645.25	2,120,386.71
Hanyang District	16,635,138.75	2,451,850.92
Hongshan District	54,850,479.25	7,368,021.45
Total	130,288,797.50	18,178,243.39

, corresponding to it, presents the specific data on the total rooftop area and PV installed capacity of each district.

The total rooftop area of the seven central urban districts in Wuhan is approximately 130 square kilometers, and the PV installed capacity is about 18.18 million kilowatts. Qingshan District appears dark red in the heat-map, indicating a relatively high level of PV installed capacity per unit area. Its total rooftop area is 10,546,692.75 square meters, and the PV installed capacity reaches 1,559,301.78 kilowatts. As an industrially concentrated area, Qingshan District has large-sized and regularly structured rooftops of industrial factories, providing ideal conditions for the large-scale installation of PV facilities. The high demand for electricity in industrial production prompts enterprises to make full use of rooftop space to install PV equipment, aiming to reduce electricity costs and achieve energy self-sufficiency. Meanwhile, compared with other districts, Qingshan District is less disturbed by geographical factors such as lakes, and its concentrated land use is conducive to the large-scale development of PV projects. Jiang’an District shows red, indicating a relatively high PV installed capacity per unit area. Its total rooftop area is 14,677,645.25 square meters, and the PV installed capacity is 2,120,386.71 kilowatts. As an important commercial and residential area in the city, Jiang’an District has a high building density and limited land resources. In this context, increasing the PV installed capacity per unit area is the key to efficiently utilizing the limited rooftop space. Qiaokou District and Jianghan District appear orange in the heat-map, with a medium-level PV installed capacity per unit area. The total rooftop area of Qiaokou District is 10,353,678.25 square meters, and the PV installed capacity is 1,465,183.39 kilowatts; the total rooftop area of Jianghan District is 8,212,744.25 square meters, and the PV installed capacity is 1,170,985.01 kilowatts. As traditional commercial areas, these two districts are densely populated with commercial buildings. The rooftops of commercial buildings often need to accommodate functions such as advertising and equipment placement, which restricts the space for PV installation. At the same time, the rooftop structure and orientation may also be unfavorable for the installation of PV panels and the improvement of power-generation efficiency. Hanyang District appears dark red. Its total rooftop area is 16,635,138.75 square meters, and the PV installed capacity is 2,451,850.92 kilowatts, with a relatively high PV installed capacity per unit area. This benefits from the existence of large-scale industrial buildings in the district, which provide good conditions for PV installation. Wuchang District and Hongshan District appear relatively light in color in the heat-map, with a low PV installed capacity per unit area. The total rooftop area of Wuchang District is 15,012,419.00 square meters, and the PV installed capacity is 2,042,514.14 kilowatts; the total rooftop area of Hongshan District is as high as 54,850,479.25 square meters, and the PV installed capacity is 7,368,021.45 kilowatts. The low PV installed capacity per unit area in these two districts is closely related to geographical conditions, in addition to building-related factors. Wuhan has numerous lakes, and Wuchang District and Hongshan District are dotted with a large number of lakes, such as East Lake. Lakes occupy large areas, reducing the land area available for building construction and, thus, making rooftop resources relatively scarce.

For areas with a high PV installed capacity per unit area, such as Qingshan District, Jiang’an District, and Hanyang District, the layout of PV systems can be further optimized to improve energy utilization efficiency, and integration with energy-storage technologies can be explored to achieve a stable energy supply. For areas with a low PV installed capacity per unit area, such as Wuchang District and Hongshan District, on the one hand, it is necessary to deeply analyze building-related restrictive factors, and through means such as policy guidance and technological innovation, tap potential PV resources. For example, subsidy policies for rooftop renovation can be formulated for old residential areas to encourage the installation of PV equipment. On the other hand, the impact of geographical conditions needs to be fully considered. In urban planning, rational layout should be carried out, and PV projects should be preferentially planned in areas not restricted by lakes. At the same time, innovative models such as floating PV power plants on water surfaces should be explored to improve the development and utilization level of PV energy. In addition, during the urban planning and construction process in each district, the development needs of PV energy should be fully considered, appropriate rooftop space should be reserved, and architectural design should be optimized to create favorable conditions for the sustainable development of the PV industry. Moreover, the promotion of PV technology can be strengthened to improve the awareness and participation enthusiasm of residents and enterprises for PV energy, thereby promoting the comprehensive development of PV energy in the central urban area of Wuhan and facilitating the city’s green transformation.

4. Conclusions

This study demonstrates the value of deep learning for city-scale rooftop PV potential assessment and offers three principal contributions. First, we constructed and publicly documented a high-resolution aerial imagery dataset of Wuhan rooftops, confirming that deep convolutional networks can accurately delineate roof areas. Second, a systematic comparison of four segmentation architectures identified U-Net with a ResNet50 backbone as the top performer, delivering ~0.94 overall accuracy and crisp roof boundaries. Third, the optimised model revealed approximately 130 km² of usable rooftop area in Wuhan’s core districts—equivalent to ~18.18 GW of installable PV capacity—thus providing quantitative guidance for distributed-solar planning.

Despite these merits, two limitations should be acknowledged. (1) Roof-type diversity: sloped or highly irregular roofs may introduce geometric-to-planar area discrepancies that marginally bias the total rooftop-area estimate. Future work will incorporate roof-tilt correction to refine area calculations. (2) Geographical vs. energy potential: our study estimates geographic (installable-area-based) potential without modeling time-varying solar irradiance and shading. Integrating high-resolution climate data and annual yield simulations (e.g., via urban climate models) will be an important next step.