Deep Learning Method for Evaluating Photovoltaic Potential of Rural Land Use Types

Abstract

Rooftop photovoltaic (PV) power generation uses building roofs to generate electricity by laying PV panels. Rural rooftops are less shaded and have a regular shape, which is favorable for laying PV panels. However, because of the relative lack of information on buildings in rural areas, there are fewer methods to assess the utilization potential of PV on rural buildings, and most studies focus on urban buildings. In addition, in rural areas, concentrated ground-mounted PV plants can be built on wastelands, hillsides, and farmlands. To facilitate the overall planning and synergistic layout of rural PV utilization, we propose a new workflow to identify different types of surfaces (including building roofs, wastelands, water surfaces, etc.) by applying a deep learning approach to count the PV potential of different surfaces in rural areas. This method can be used to estimate the spatial distribution of rural PV development potential from publicly available satellite images. In this paper, 10 km² of land in Wuhan is used as an example. The results show that the total PV potential in the study area could reach 198.02 GWh/year, including 4.69 GWh/year for BIPV, 159.91 GWh/year for FSPV, and 33.43 GWh/year for LSPV. Considering the development cost of different land types, several timespans (such as short-, medium-, and long-term) of PV development plans for rural areas can be considered. The method and results provide tools and data for the assessment of PV potential in rural areas and can be used as a reference for the development of village master plans and PV development plans.

1. Introduction

1.1. Background

According to the International Energy Agency (IEA) [1], which released the Global Energy Review: Carbon Emissions 2021, global CO₂ emissions from energy combustion and industrial processes reached 36.3 billion tons in 2021 due to the economic recovery since the COVID-19 pandemic, which is up by 6% year-on-year from 2020 and has reached the highest annual level ever. The world is not fully on the right path to a sustainable recovery and needs to pay more attention to the development of renewable energy sources. Solar energy is considered one of the best alternatives to traditional energy sources, and the proper use of solar energy is of great significance in alleviating energy depletion as well as curbing environmental degradation and other problems. Photovoltaic power generation, which refers to the conversion of solar radiation into electricity by using components such as panels, reached a cumulative installed capacity of more than 942 GW by 2021. The contribution of photovoltaic power generation accounts for an average of 5% of energy use in the world. Compared with wind power and hydroelectric power generation technologies, solar power generation is the renewable energy generation technology with the highest potential for large-scale development and application, as it has a broad market demand and good prospects for development [2].

As a result, China’s solar energy development has gained momentum in the past decade. “In 2021, China’s new PV installation scale reached 54.88 GW, accounting for 31% of the global” [1]. China’s photovoltaic development in rural areas has been very rapid and will therefore become an important area of China’s solar photovoltaic development in future. As rural living standards continue to improve, electricity consumption in rural areas is increasing as photovoltaic utilization technology becomes increasingly mature and construction costs are more easily accepted by rural households, which provides favorable conditions for the development and utilization of rooftop solar energy in rural areas. In China, cities are dominated by high-density buildings due to concentrated urban planning and construction. The area of buildings in rural areas is much larger than in urban areas, the proportion of single buildings is higher, and shading between buildings is less. In addition, there are also large surfaces on barren hills and ponds in rural areas that are ideal for the development of solar PV utilization systems.

However, it is necessary to assess solar potential and guide the planning and development of solar resources for rural areas, which will facilitate the rational and efficient use of solar energy and reduce the waste of resources as well as carbon emissions.

1.2. Overview

Thanks to the available information on urban building models and solar radiation, most studies have focused on the analysis of solar photovoltaic energy in urban areas but fewer on rural areas.

Because of buildings mainly concentrating in urban spaces and solar energy use mainly being based on the surfaces of buildings, studies of solar energy potential in the urban sector area have mainly focused on buildings [3], from individual buildings [4] and building groups [5] to urban form [6]. In terms of applications, there are also many practical examples of urban solar planning [7,8,9]. It can be seen that the research on urban solar potential has become more mature and formed a more complete system.

For rural areas, the overall impact of building interactions and spatial patterns on solar potential is limited due to the low height and relative dispersion of buildings. Although solar potential from building surfaces also cannot be a major source of rural solar energy use, relevant building-specific studies can provide sufficient references for solar energy research and assessment in rural areas.

By combing through the relevant research work, we found that the methods used to quickly assess the solar potential of certain areas of cities are divided into three main types, which are shown in

Table 1. Methods for accessing to the distribution of urban building roof.

Common Methods	Data Required	Assessment Scale	Calculation Speed	Calculation Accuracy	Research Heat
Empirical coefficient method	Statistical data such as roof surface ratio per capita population	City-level, national-level PV generation potential analysis	High calculation speed	Low precision	Due to the accuracy problem, few publications have been written in recent years
3D model method	High-resolution LiDar or DSM data	City-level, community-level generation potential analysis	Low calculation speed	High precision for building façade and roof	Most of the studies have always been created with the help of national research data
Deep learning method	Easily obtainable free satellite images	Applicable to a flexible scale of PV generation potential analysis	High calculation speed	High precision for building roofs, but not façade information	It has arisen in recent years and is widely used in cities, but rarely in rural areas

.

The first category uses some sampling or research to obtain empirical ratios such as a method based on determining the ratio of roof surface per capita. The total roof surface is then calculated by multiplying this ratio by the total population of the target area [10]. Typical studies include those by L.K. Wiginton [11] and Mainzer et al. [12].

The second category is based on 3D building models, and most traditional building roof-area estimates are based on building information generated by LiDAR or obtained from known urban digital surface models (DSMs). For example, Julieta [10] obtained the total roof area of one Spanish region based on information from the land registry department and classified land use to estimate solar resource utilization potential in Spain. Other similar studies include David A. Jacques et al. [13] and Hong et al. [14].

The third category is based on deep learning methods, which enable large-scale analysis and evaluation to be carried out by using images. Semantic segmentation network models based on deep learning can train end-to-end networks on image inputs of arbitrary resolution size, reduce practice while processing large amounts of data, and replace a large amount of repetitive manual work. This method has been wildly applied in study of land use classification by many scholars such as Ce Zhang [15], Fisher [16], and Pan [17]. They have used deep learning techniques to quickly classify satellite images to filter out usable areas. Some scholars have also used it in predictive studies of urban solar power. Huang et al. [18] used a U-Net network to predict solar power generation in Wuhan and made recommendations for development. Zhang Chen et al. [19] made predictions of PV potential for different types of urban land by CNN network.

In fact, deep learning-based approaches will largely help rural areas to carry out solar energy analysis and evaluation of buildings despite the lack of basic data support in rural areas.

In addition to building rooftops, there are many other land types in rural areas, such as ponds and wastelands, where solar panels can be installed. This is a peculiarity of rural areas and some studies of solar potential assessment provide references for it.

For example, Coruhlu et al. [20] assessed the solar potential of some areas in Turkey by using GIS tools and the AHP (analytical hierarchy process) method. Amaducci et al. [21] studied the installation of photovoltaic equipment on corn farms by using a solar tracking system named Agrovoltaico. Other, similar studies, such as those by Doljak et al. [22] and Yu et al. [23], focusing on PV plant-siting evaluation can also provide meaningful insights.

The aim of this paper is to create a comprehensive measurement of the PV potential of all land types in rural areas with the help of new methods and tools for deep learning. This measurement is used to quickly analyze and assess the relationship between solar potential and land use in rural areas. It enables planning authorities to understand the distribution of PV potential in rural areas and provides them with tools and data to support the development of rural master plans and PV development plans. In addition, international scholars can use the concepts of this study to examine the synergy between solar potential and land use planning in rural areas around the world, which could be the basis for a rational development plan.

2. Data and Methods

A three-stage method was applied in this study as shown in Figure 1. The first stage was to carry out an assessment of areas suitable for solar energy utilizations, in which land use data and elevation DEM data obtained from open access were used. The second stage, in which a tool based on a U-net convolutional neural network was applied to help carry out partial calculations, was to calculate the value of the area suitable for PV installation based on the assessment in the previous stage. The last stage was to calculate PV utilization potential based on suitable area, in which different PV conversion efficiencies and area conversion factors were applied for each type of land suitable for PV installation.

2.1. Study Area

The study area involved in this research is located in Jiangxia District, Wuhan, Hubei Province, China, which is one of the six distant urban areas administrated by Wuhan, with many rural settlements and abundant rivers and lakes. Therefore, the exploitation of local solar energy resources to achieve partial energy self-sufficiency can, to a certain extent, effectively reduce the overall energy consumption of Wuhan City and reduce the investment and construction costs of electric power infrastructure. In addition, Jiangxia has carried out some practical work in the field of solar energy utilization, such as the construction of a zero-energy town near Liangzi Lake which uses photovoltaic technology [24] and the 350 kwP distributed photovoltaic project in the Canglong Island area [25], so it can be said that the local area has some of the basic conditions for the development and utilization of solar energy.

Therefore, part of Jinkou Street, which is a rectangular area of 100 square kilometers (10 km × 10 km) in Jiangixa District, was selected as the study area (Figure 2).

2.2. Data Resources

In this study, two main types of data resources were applied which are as follows: (1) Google satellite images; (2) land use data.

2.2.1. Google Satellite Images

Google high-resolution satellite images were culled from public sources. The Google Maps platform provides an open Map Static API to obtain urban satellite images, which can return images through URL requests and can formulate the location and size of images, zoom levels and types of maps. At the image level, there are 22 image-zoom levels for Google Maps data, and each zoom level corresponds to a different resolution.

In this study, the satellite image data of 2020 for study area with 17 levels of image zoom levels were used, as shown in Figure 3.

2.2.2. Land Use Data

Land use data is data information reflecting the state, characteristics, dynamic changes, distribution characteristics of land use systems, and land use elements, as well as human development and utilization of land, governance and transformation, management and protection, and land use planning.

In this study, the 2020 Chinese national land use data obtained from open sources (Chinese Resource and Environment Science and Data Center (https://www.resdc.cn/, accessed on 15 February 2023)), which is a digital image based on Landsat TM, is utilized, and a fully digital human–computer interaction remote sensing fast extraction method is used.

The land classification in the dataset refers to the existing land use/land cover classification system in China, and the land types are divided into 6 primary classifications and 22 secondary classifications, as shown in

Table 2. The land use types of 4 broad categories and 22 subcategories.

Broad Category	Subcategory
Cultivated land	Paddy filed
	Dryland
Forestlands	Closed forestlands
	Open forestlands
	Shrublands
	Other forestlands
Grasslands	High-coverage grassland
	Medium-coverage grassland
	Low-coverage grassland
Waters	Rivers and ditches
	Lakes
	Reservoirs and ponds
	Snow/ice
	Mudflats
Urban, village and mining lands	Urban and towns
	Villages
	Other construction land
Unused land	Sandy areas
	Gobi Desert
	Saline land
	Marshlands Bare exposed land
	Bare exposed rock or gravel

.

2.3. Land Types Suitable for Solar Energy Utilization

This section will focus on the land types suitable for solar energy utilization and the appropriate utilization methods.

2.3.1. Rural Areas

For villages, the main way to utilize solar energy is to install photovoltaic equipment on the roofs of buildings to generate electricity or to heat water. Depending on the shape of the building’s roof, photovoltaic use can be divided into two forms: flat roof and inclined roof; additionally, according to the installation form of photovoltaic modules, it can be divided into horizontal installation and inclined installation.

In this research, the land type of rural areas is “Urban and towns” (only small towns in research region) and “Villages” (in Section 2.2.2), and in order to simplify the research and calculations, only the condition of “flat roof, horizontal installation” is considered.

2.3.2. Waters

In recent years, the deployment of floating solar photovoltaic system (FSPV) on water surfaces has become a new direction for research and practice in the field of solar energy utilization, especially for areas with relatively abundant rivers and lakes. For rural areas with abundant water resources, an FSPV system can also avoid the use of valuable agricultural land and reduce deforestation while the PV panels can use the water for cooling.

FSPV systems provide an efficient solution for energy production without significant demands on water and land resources. This new solar PV technology involves the use of floating structures to install PV panels on water surfaces. A typical floating PV equipment is shown in Figure 4. In addition to rural areas, the installation of FSPV devices can also alleviate regional energy demand, especially in less developed areas. Thus, the available water space in rural areas is included in the PV potential assessment in this study, and the land type is “Reservoirs and ponds” in Section 2.2.2.

2.3.3. Unused Lands

Unused land in rural areas is another area with potential for solar energy utilization, and the main way to utilize it is by installing large-scale photovoltaic (LSPV) systems, which are a type of solar energy utilization that requires a large amount of land. Compared with urban areas, rural areas are in a better position to install LSPV systems. In China, the construction of LSPV projects should occupy little or no arable land as well as occupying environmentally friendly areas in order to protect the ecological environment [23]. Therefore, the sites where LSPV systems can be installed in rural areas are limited, such as sand, saline, swamps, bare land, etc. In addition, some low-coverage grasslands are also suitable for the installation of photovoltaic systems via a composite development and utilization approach. Meanwhile, the installation of LSPV equipment needs to take the influence of regional slope into account, and the slope condition of suitable land is set to 2.1% in this study based on relevant studies [28].

The land types of unused land contain all subcategories in “Unused land” and “Low-coverage grassland” in Section 2.2.2.

2.4. U-Net Neural Networks

In this study, the main role of the U-net neural network is to build tools which can calculate solar energy utilization potential by identifying roofs in rural sites.

The network consists of two main components (Figure 5): feature extraction (similar to the classical VGG network) and upsampling. The feature extraction part consists of four iterations of convolution and maximum pooling operations. Each time an image passes through the pooling layer, a scale is added, resulting in a total of 5 scales, including the original image. The repeated convolution operation in the upsampling part is to reduce the number of feature channels by half. Each time the network performs upsampling, the obtained feature map is stitched together with the downsampled part of the feature map of the same size. In the last layer, a 1 × 1 convolution and an S-type activation function are used to obtain a probabilistic heat map. Thus, the network contains four pooling layers and four upsampling layers [29,30].

The scripting language Python was used to implement the neural network in this study and was written in the PyCharm environment. The deep learning of the program consists of five modules: dataset loader, neural network, loss function, trainer, and tester. The method can achieve an accuracy of about 85% in recognizing roofs. The results of roof recognition in the research region are as follows (Figure 6). The white area represents the building roof area identified by the deep learning tool, and the black area represents other non-building property areas.

2.5. Calculation Method for Solar Energy Utilization Potential

The PV utilization potential involved in this study was calculated using a method commonly used in this field [31]. The calculation method takes three main aspects into account: (1) the solar radiation received on the surface of the PV module, (2) the effective area of the PV module, and (3) the efficiency of the PV module. The calculation method follows the following equation for the calculation.

$${PV}_{Total}=R_{Total}/{10}^6\times S_{model}\times {\eta }_{model}$$

(1)

In this formula, ${PV}_{Total}$ represents the annual PV power potential in MWh (GWh/year); $R_{Total}$ represents the solar radiation received per unit area of the PV module surface, i.e., the total horizontal radiation, in kWh/m²⋅year; ${\eta }_{model}$ represents the overall efficiency of the PV module and PV system.

The parameter $R_{Total}$ in Equation (1) is one of the important parameters for calculating radiation on the surface of PV modules, including both direct solar radiation (DNI, diffuse normal irradiation) and scattered radiation (DIF, diffuse horizontal irradiation). The total horizontal radiation data used in this study are obtained from the Global Solar Atlas global solar resource dataset published by the World Bank [32], which is part of the Global Renewable Energy Mapping Program and can provide users with estimates of the solar resource potential of their regions on a global scale. This solar resource dataset has been produced with consideration of the influence of meteorological and climatic factors in different regions and cities and can visually reflect the strong and weak distribution of solar resources in different regions and cities. The total horizontal radiation in Wuhan used in this study is 1241 kWh/m²⋅year.

For the parameters $S_{model}$ and ${\eta }_{model}$ in Equation (1), since this study involves different land use types and PV utilization methods, the above two parameters need to be defined further depending on the different site types within the study area, which means that the above two parameters can be further defined thus: (1) $S_{model\left(R\right)}$ and ${\eta }_{model\left(R\right)}$, which are applicable to rural areas; (2) $S_{model\left(W\right)}$ and ${\eta }_{model\left(W\right)}$, which are applicable to water; (3) $S_{model\left(L\right)}$ and ${\eta }_{model\left(L\right)}$, which are applicable to the unused land.

2.5.1. $S_{model\left(R\right)}$ and ${\eta }_{model\left(R\right)}$

In rural areas, the main source of PV potential is roof-mounted distributed PV modules, and their $S_{model\left(R\right)}$ value is mainly derived from the roof of the building in scope, while taking into consideration that the building roof cannot be fully installed with PV equipment, so it can be obtained by means of the roof area of the building in scope and the roof availability coefficient, which is the following formula.

$$S_{model\left(R\right)}=S_{roofs}\times {\eta }_{roofs}$$

(2)

In this formula, $S_{roofs}$ is the area of building roofs, and ${\eta }_{roofs}$ is the roof availability factor.

Many studies show that the building roof availability factor is maintained at 40%, so ${\eta }_{roofs}$ is set to 40% in this paper. $S_{roofs}$ is then obtained by recognizing satellite images through the U-net neural network tool, and the ${\eta }_{model\left(R\right)}$ is set to 85%.

2.5.2. $S_{model\left(W\right)}$ and ${\eta }_{model\left(W\right)}$

The land class is mainly bodies of water such as lakes and reservoir pits. $S_{model\left(W\right)}$ is mainly derived from water surface area, and again, since not all water surfaces can be utilized, they need to be converted via the utilization factor, i.e., the following equation.

$$S_{model\left(W\right)}=S_{water}\times {\eta }_{water}$$

(3)

In this equation, $S_{water}$ is the water surface area and ${\eta }_{water}$ is the water surface availability factor.

In many FSPV studies, the utilized area of the water surface is maintained between 10% and 40%. In this study, ${\eta }_{water}$ was set to 10%, while the ${\eta }_{model\left(W\right)}$ was set to 16.92% (conversion efficiency is 18% and system efficiency is 94%) [33,34,35,36,37].

2.5.3. $S_{model\left(L\right)}$ and ${\eta }_{model\left(L\right)}$

This parameter is applied to assess the PV potential of rural unused land whose main use is occupation by LSPV equipment for electricity generation. Again, not all sites can be fitted with PV equipment and therefore need to be discounted by a factor calculated as follows:

$$S_{model\left(L\right)}=S_{Land}\times {\eta }_{Land}$$

(4)

In the formula, $S_{Land}$ is the area of unused land, and ${\eta }_{Land}$ is the land availability coefficient.

In many studies, ${\eta }_{Land}$ was usually in the range of 35% to 70%, as shown in

Table 3. Land use coefficients in studies.

Location	${\mathit{\eta }}_{\mathit{L}\mathit{a}\mathit{n}\mathit{d}}$
India	35%
Bangladesh	7.86%
China	70%
Sultanate of Oman	59.85% and 25.39%

[38,39,40,41].

In this study, ${\eta }_{Land}$ is set to 35%, the conversion efficiency of the PV device $S_{Land}$ is set to 12.75% (conversion efficiency is 15% and system efficiency is 85%).

3. Results

3.1. Land Use in the Study Area

The results of the land types within the study area are shown in Figure 7.

As the figure shows, the top three land types in terms of percentage are Paddy-filled, Open forestlands and Lakes. The proportion of their area to the total study area was 20.10%, 15.60%, and 13.86%. And the area of land type “Urban and towns” as well as “Villages” is 7.96 km² and 44.42 km², with the land use percentage 0.80% and 4.44%. The occupancy of each land type in the study area is shown in Figure 8.

Further, the identification of building roofs with U-net neural network tools was carried out for land type “Urban and towns” and “Villages”. The results show that the total building roof area in the study region is 74,062.45 m².

3.2. PV Potential in the Study Area

The acquired data regarding land area and roof area have been applied to carry out PV potential assessment for the types of land suitable for solar energy utilization. An image of the distribution of PV potential in study area has been carried out as shown in Figure 9. Meanwhile, the quantified results show (

Table 4. The land use types of 4 broad categories and 22 subcategories.

PV Utilization Types	Applicable Land Types	PV Potential (GWh/year)	Percentage
BIPV	Rural areas	4.69	2.37%
FSPV	Water	159.91	83.75%
LSPV	Unused land	33.43	168.88%
Total	——	198.02	100.00%

) that the total value of PV potential in study area can reach 198.02 GWh/year, where the PV potential from BIPV is 4.69 GWh/year, from FSPV is 159.91 GWh/year, and from LSPV is 33.43 GWh/year. It can be seen that the largest contribution to the PV potential of the study region is made by FSPV.

Figure 10 further shows the distribution of potential for each PV utilization type (BIPV, FSPV, and LSPV). It can be seen that the PV potential of BIPV comes mainly from the concentration of large rural settlements, the PV potential of FSPV is from large lakes and some ponds, and the PV potential of LSPV is from a large area of unused land.

3.3. PV Potential Based on Land Types

Further, we analyzed the PV utilization potential for different land type conditions, including the distribution of PV potential for each land type sample and the PV utilization potential per unit land area condition.

For rural areas, Figure 11 shows the PV potential of samples of rural areas. It can be seen that the highest PV utilization potential for rural areas is 0.062 GWh/year and the average PV utilization potential is 42.25 kWh/year. The PV potential output per unit area of land can reach 0.09 GWh/km²·year.

For water samples (Figure 12), the PV utilization potential can reach a maximum of 4.18 GWh/year, but with an average value of 0.014 GWh/year. The PV potential per unit can reach 2.48 GWh/km²·year.

As well as unused land samples (Figure 13), the highest PV potential can reach 6.88 GWh/year, and the average PV potential 0.003 GWh/year. The PV potential per unit can reach 56.66 GWh/km²·year.

A further comparison of PV potential per unit area for three land types (

Table 5. PV potential per unit area for land types.

Land Types	Type	Area (km2)	PV Potential (GWh/year)	PV Potential per Unit Area (GWh/km2·year)
Rural areas	BIPV	92.68	4.69	0.05
Water	FSPV	64.52	159.91	2.48
Unused land	LSPV	0.59	33.43	56.66

) shows that the highest PV potential per unit area is the value of “Unused lands” and the lowest value is that of “Rural areas”. The main reason for this may be that a higher percentage of open land can be fitted with PV panels, so more electricity can be produced per unit area of land, while the main source of PV in rural residential sites is roofing PV panels. However, not all of the range has a large number of buildings and not all roofs can have PV panels installed. Therefore, after discounting these factors, the actual area available for power generation is less, which affects the result of PV potential per unit area.

4. Discussion

In the methods and results of this research, there are several points worth discussion.

This study is one of the studies that assess solar PV potential in rural areas. Compared with other studies, this study aims to achieve a comprehensive assessment of solar energy utilization potential in rural areas by constructing a research path integrating the PV potential of rural buildings and other land PV utilization.

Free public satellite images were used in this study, which can effectively solve the problem of missing geographic data in rural areas. It is worth stating that since satellite images do not contain height information, they cannot be used to analyze the shading between adjacent buildings, but most rural buildings are single-story buildings with relatively low floor area ratios and building heights, so the mutual shading effect between buildings is small, and some related studies have also proved that this method is more reliable for rural areas. For other sites, we have also fully considered the technical conditions and analyzed the feasibility of installing different PV installations for different site conditions, instead of assessing rural land only at the geographic potential scale.

This study allows the relationship between solar PV utilization potential and different rural lands to be assessed in order to determine what kinds of rural terrain are suitable for solar energy development. It will contribute to the development of solar potential in rural areas, thereby reducing the cost of building and operating regional power infrastructure and enabling localized power supply for industrial development to be provided in rural areas. In addition, the authors would like to point out that no analysis of the economics of solar energy development has been carried out in this paper due to the cost and price differences between different geographical areas.

Some experts having mentioned the impact of solar panel coverage on ecological systems since the natural environment in Wuhan is unique due to its large water area. Due to the limitation of the installation factor, not all water surfaces can be used for PV installation, and we also considered the fact that large rivers and lakes are not suitable for PV installation to avoid damage to the existing ecological environment. In addition, the theoretical power generation capacity generated by using only smaller water surfaces is already considerable, and combined with the PV power generation capacity of other installation methods, it can meet part of the production and living electricity demand in rural areas. At the same time, before construction in Wuhan, the project will require environmental image assessments and permission from the environmental authorities. Moreover, our study is intended to apply only to early planning predictions; some technical tools can be subsequently used to reduce the impact on the environment. Thus, it can be assumed that the influence of solar panel coverage on the natural environment will be limited.

5. Conclusions

In this study, a new method and workflow based on a deep learning approach is constructed for rapid assessment of the solar energy potential of different types of land in rural China. This method helps to clarify the distribution of solar potential in rural areas and suitable areas for development before carrying out solar energy planning in rural areas. This improves the scientific and implementable nature of solar energy planning.

(1) This study discusses all the installable surfaces of PV in rural areas. Specifically, these include: building roofs, water surfaces, and wastelands. Then, we build the dataset by using public, free-to-use satellite images as the data source and perform correlation validation using the classical U-Net convolutional neural network as a basis.

(2) In rural areas, more circumstances have to be considered for better utilization of solar energy. Using the study area of this paper as an example, the total value of PV potential in rural areas can reach 198.02 GWh/year, of which BIPV is 4.69 GWh/year, FSPV PV potential is 159.91 GWh/year, LSPV PV potential value is 33.43 GWh/year. In rural areas, the share of building roofs is relatively low, only 2.37%.