Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Communications Earth & Environment volume 6, Article number: 1006 (2025) Cite this article
Growth in solar photovoltaic capacity supports grid decarbonization but can result in land transformation. Quantifying land–solar interactions is hampered by inconsistent methods and data. We develop a consistent, replicable framework to quantify land-solar interactions and apply it to annotated aerial imagery covering 719 solar photovoltaic projects (13,272 megawatts of installed capacity) connected to the Western Interconnection in the United States. We train a deep-learning convolutional neural network to characterize solar photovoltaic land footprints, post-process outputs with geospatial land-cover overlays, and compute land-use efficiency and energy-normalized land transformation per project. Across the sample, mean capacity-based land-use efficiency is 24.7 ± 15.2 watts per square meter and mean lifetime land transformation is 0.846 ± 0.722 square meters per megawatt-hour; regional differences and engineering choices explain project-level variability. Our open-source inventory and method enable more consistent large-scale assessments of planning, life cycle impacts, and ecological trade-offs of solar expansion.
Renewable energy sources are an important conduit through which energy-related greenhouse gas emissions can be reduced, particularly photovoltaics (PV) due to the abundant and ubiquitous nature of solar energy^1,2,3. Despite having only generated 3.2% (856 TWh) of global electricity generation in 2020, the PV industry has maintained high growth rates for the past decade and its cumulative capacity may reach ten terawatts by 2030^4,5. The growth of ground-mounted PV may experience land constraints, yet the degree to which it may be constrained has yet been confirmed due to inconsistent methodologies and a lack of comprehensive datasets. Further, ambitious solar energy development could potentially create unanticipated knock-on effects on land use on a global scale if existing land uses are displaced, highlighting the importance of optimizing co-use of land. As the world transforms its energy system toward lower carbon and renewable sources, understanding not only the benefits but also how to overcome potential drawbacks of solar PV has important implications for sustainability. Our analysis contributes a transparent and systematic method to quantify land-solar interactions and comprehensive dataset of to be scientifically accurate 719 solar power PV projects in the U.S. portion of the Western Interconnection.
The installation of solar PV projects captures free and abundant solar irradiance over large parcels of land which are neither free nor unlimited. The issue of land use has been perceived as a limitation to the large-scale growth of ground-mounted solar power, with early studies suggesting that it can compromise the overall sustainability and return on investment of solar energy systems^6,7,8. A total of 32 environmental impacts have been linked to the utilization of solar energy, indicating that high growth in solar power may result in unintended consequences to various environmental impact categories⁹. At the same time, solar energy has been noted as having both lower and higher land use in comparison to wind and fossil fuel powered electricity using life cycle assessment (LCA), a cradle-to-grave analysis of the impacts of products and processes^10,11. Land transformation (square meters per megawatt hour ((frac{{m}^{2}}{{MWh}}))) serves as an important baseline metric for LCA in terms of area per unit of electricity generated. The quantification of land-solar interactions can enable an understanding of how to ameliorate impacts while supporting techno-ecological synergies^{6,12,13,14,15,16,17}; this framework to engineer solutions that benefit both solar technologies and ecological systems by design has become a major thrust of research in recent years. Transforming our perspective on systems design can ensure that solar energy is developed to align with both climate and sustainable development goals (SDGs)¹⁸.
Beyond LCA, inconsistent metrics, data, study boundaries, and methods have produced conflicting results for land-energy interactions^6,19,20,21. Over short timescales (e.g., <10 years), renewable energy systems have been noted as much more land-intensive than fossil-fuel-based systems^10,22,23,24. Solar power has also been noted as having lower scores in capacity-based land use efficiency (watts of installed capacity per square meter ((frac{{W}_{c}}{{m}^{2}}))) and generation-based land use efficiency (watts of operational capacity per square meter ((frac{{W}_{o}}{{m}^{2}})))¹¹ in comparison to non-renewable alternatives with strong dependence on geographical locations²⁵. Such results overlook the fact that renewable power can increase its cumulative electricity produced without expanding its land footprint whereas non-renewable power requires the continual development of new land for fuel extraction²⁶. Furthermore, solar PV facilities can increase their land use efficiency by having higher packing factor—the ratio of PV array area to total site area for a facility²⁷. Considering the implicit time factor in land-use metrics, renewable energy sources can have comparable or even lower land impacts when compared to fossil counterparts when examined over longer timescales, such as the average lifetime of solar projects (30 years)^10,24,28. Solar power can reach land equivalency (i.e., the amount of time it takes for fossil-fuel-based systems and renewable power systems to have the equivalent land footprint and cumulative energy production) with natural gas and coal alternatives in less than a decade^10,24.
Systematic research that quantifies the land-solar interactions  covering large populations of PV projects across extensive tracts of land remains relatively limited. Existing analyses rely on relatively small datasets (e.g., <200 power plants in Hernandez et al.⁶) or present results focused on the direct panel area rather than the total project site^29,30,31 (i.e., including the panels and all land use up to the project fence line), limiting the ability to directly use findings to confirm the land-use impacts associated with current and future solar energy generation²⁴. Without readily available data describing the land use of solar PV facilities, the potential higher-order, indirect effects of solar land use on the global scale become even more difficult to measure. Furthermore, the variable use of metrics in existing literature prevents broad consensus about land-solar interactions, which can be circumvented by adopting a transparent methodology and consistent set of metrics²¹. Our analysis of the 719 U.S. solar PV projects (Fig. 1) contributes a systematic and replicable methodology that produced a dataset quantifying  the resulting land transformation (the land area altered by the solar facility per unit of energy produced during its 30-year lifetime) and land-use efficiency (the land area altered per unit of installed nameplate capacity or electricity generation). The power plants in our sample are all connected to the U.S. Western Interconnection, which occupies an area rich in solar energy in terms of global horizontal irradiance (GHI in ({kWh}/frac{{m}^{2}}{{year}}))—a close proxy for the total endowment of solar resources available at each given location on the earth’s surface.
These figures show the locations of solar PV arrays documented for this study (Energy Technology and Policy Assessment Research Group, ETAPA) and USGS studies. The 2016 U.S. Geological Survey (USGS) data identifies the locations and footprint of 740 solar PV facilities (1207 clusters of PV arrays) for the conterminous U.S. by 2015 (green). We document 719 solar PV facilities (8089 clusters of PV arrays) in the U.S. Western Interconnection (separated by the blue stroke) by 2019 (blue). The base layer shows the annual average GHI produced by the National Renewable Energy Laboratory (NREL). a This map shows the locations of solar PV arrays documented in the USGS data that are connected to the Western Interconnection. b This map shows the locations of solar PV arrays examined in our study that are connected to the Western Interconnection.
By combining deep learning and energy systems analysis, we present a workflow that can be broadly applied to solve inconsistencies in the quantification of land-solar interactions while elucidating the use of the built environment. While previous studies have documented efforts to localize and create inventories of solar PV installations in the U.S.^{29,31,32,33,34,35,36,37}, access to the resulting data has been limited^33,34,36 and the annotation of solar PV installations has been focused on the panel arrays²⁹). We introduce a replicable and efficient machine-learning-based approach for analyzing land-use associated with solar PV that differentiates between the panel array area and the full project area of solar PV facilities. Our approach is applicable across diverse regions and installation types, not only generating robust datasets but also providing consistent metrics and transparent, replicable methods that offer critical insights for resolving inconsistencies in LCA and other macro-scale energy analyses.
We examined 719 utility-scale solar PV facilities from U.S. Energy Information Administration (EIA) records that are connected to the Western Interconnection (Fig. 1a). Each of our utility-scale facilities has a nameplate capacity of at least 1 megawatt as defined by the International Energy Agency (IEA)³⁸, with the median, mean, and total nameplate capacity being 3.5 megawatts, 19.7 megawatts, and 13,272 megawatts respectively. Our dataset also includes 110 utility-scale rooftop PV installations, predominantly situated on substantial structures such as university campuses, government laboratories, and corporate facilities, averaging 2.1 megawatts in nameplate capacity and comprising 1.8% (236.3 megawatts) of the total analyzed capacity. Our results suggest that solar projects examined in the study area cover 538 ({{{mathrm{km}}}}^{2}) of land, 416 ({{{mathrm{km}}}}^{2}) of which is covered by solar panels. Our PV panel footprint results show strong alignment with prior works completed by the U.S. Geological Survey (USGS) and Lawrence Berkeley National Laboratory (USPVDB)^30,31 (Fig. 2). Out of the 719 solar PV facilities examined in our study area, 252 and 609 were once examined by the aforementioned studies (Fig. 1b). When comparing across the same PV projects, our estimation for the area of PV panel arrays is close to that of the existing data. The minor differences can be attributed to three factors: 1) the later expansion in capacity for some projects after 2015; 2) inconsistent geopositioning of specific projects; and 3) different methods in accounting for the gap separating PV panel arrays. Specifically, solar arrays separated by more than 30 meters are annotated as separate polygons in the 2016 USGS data while arrays are examined on a case-by-case basis using our approach (see methods). Differences in methods point to the need for judicious use of results from different studies as they may not have been produced with commensurable methods. Our approach provides the land-use footprint of not only the panel area but also that of the entire solar PV facility by evidence of fencing, which is largely missing from existing data^29,30,31,32. This additional information allows us to evaluate PV array spacing using the packing factor metric—the ratio of panel array footprint to that of the entire facility²⁷. An evaluation of the 719 solar PV facilities shows a median and capacity-weighted average packing factor of 74% and 75%, respectively.
The chart shows the distribution of land-use footprint estimates for the same clusters of PV panels in our findings (green), the 2016 USGS study (blue), and 2023 USPVDB (blue). The results closely mirror each other, with minor differences stemming from the different approaches toward annotation. We also contribute by providing statistics on the land-use of entire solar PV facilities. Mean values are shown as black-centered lines. Vertical error bars denote 95% confidence intervals of the mean. Y-axis is restricted to 0–2 km² for visual legibility; full distributions of project footprint reach an upper limit of 11.2 ({{{mathrm{km}}}}^{2}).
The land transformation for solar PV projects in U.S. states is, of course, influenced by the available solar resources (Fig. 3a, b). Specifically, the average land transformation of PV facilities is lower by 16% in the U.S. Southwest (Arizona, California, Colorado, Nevada, New Mexico, Texas, and Utah) than in the Northwest (Idaho, Montana, Oregon, Washington, and Wyoming). When evaluated in terms of capacity- and generation-based land use efficiency, Southwestern PV facilities outperform those located in the Northwest by 22% and 36%, respectively (Fig. 3c–f). The PV projects in Arizona and El Paso, Texas on average are the best performers among all projects in our sample. For both land transformation and land use efficiency metrics, evaluating PV projects solely on the footprint of solar arrays yields better performance results than estimates based on the total project footprint, which has important implications for future studies on PV power considering the balance of system for PV projects is increasingly likely to include potentially land-consuming subjects such as utility-scale battery packs. On average, the 609 ground-based solar PV facilities have a capacity- and generation-based land use efficiency of 24.7 (frac{W}{{m}^{2}}) and 5.8 (frac{W}{{m}^{2}}). Accounting for the varying installed capacity of PV projects, our results show a capacity- and generation-weighted average land use efficiency of 25.4 (frac{W}{{m}^{2}}) and 6.9 (frac{W}{{m}^{2}}), updating prior estimates²⁷. The average lifetime land transformation for the examined facilities is 0.85 (frac{{m}^{2}}{{MWh}}). Rooftop installations are excluded from these land-use calculations derived from project footprints, due to the absence of clearly and exclusively defined project site areas for these installations but are available in the published dataset.
These graphs show the average land transformation and land use efficiency of PV projects in each state at the panel- and project-level. The land transformation of each PV plant is calculated using the function specified in the Method & Data section. Annual generation data for each plant in 2019 is retrieved from form EIA-923. All states are connected to the Western Interconnection (WECC). The states of Arizona (AZ), California (CA), Colorado (CO), Nevada (NV), New Mexico (NM), Texas (TX), and Utah (UT) are categorized as ‘southwestern states’ (SW). The remaining states of Idaho (ID), Montana (MT), Oregon (OR), Washington (WA), and Wyoming (WY) are categorized as ‘northwestern states’ (NW). a This graph shows the average land transformation of PV projects in each state. b This graph shows the average land transformation of PV projects in each region. c This graph shows the average capacity-based land use efficiency of PV projects in each state. d This graph shows the average generation-based land use efficiency of PV projects in each state. e This graph shows the average capacity-based land use efficiency of PV projects in each region. f This graph shows the average generation-based land use efficiency of PV projects in each region.
Results differ between PV facilities using different mounting systems. On average, PV panel arrays on dual-axis tracking systems outperform their fixed-rack and single-axis neighbors within the same state by 25% and 39%, respectively in terms of land transformation (Fig. 4a). However, in terms of total land transformation for the entire project site, PV facilities built with dual-axis tracking systems are greater on average than their single-axis and fixed-rack counterparts in the same state by 58% and 13%, respectively (Fig. 4a), meaning the former has a larger land footprint than the latter. The same trend persists when evaluated using land use efficiency (Fig. 4b, c). This dichotomy arises because fixed-rack/single-axis PV panel arrays are built in a much more compact formation, taking up less total land than dual-axis panel arrays which require each individual panel to freely and fully track the movement of the sun. On the other hand, individual PV panels on dual-axis mounting systems exhibit greater efficiency in capturing sunlight during the day than their counterparts do.
These box and whisker plots show the difference in the land transformation and land use efficiency between PV projects using dual-, fixed-, and single-axis panel mounting systems. Out of the 719 PV projects examined in our study area, three are built with a mixture of dual-axis and fixed/single-axis mounting systems, 13 are built with dual-axis mounting systems, 110 are built with fixed-rack/single-axis mounting systems on rooftop, and 593 are built with fixed-rack/single-axis mounting systems on open ground. Among the 13 projects with dual-axis mounting systems, 11 are in California, two are in New Mexico and the remaining one is in Colorado. For consistent comparisons of all three mounting systems, only these three states are shown. a This graph shows the average land transformation of PV projects with different mounting systems. b This graph shows the average capacity-based land use efficiency of PV projects with different mounting systems. c This graph shows the average generation-based land use efficiency of PV projects with different mounting systems. The box represents the interquartile range, the line inside the box is the median, and the whiskers extend to the most extreme data point that is within 1.5 times the interquartile range.
Land sparing—including siting generation on already-developed surfaces such as rooftops or redeveloped brownfields—can support the potential synergy between technology and ecological health for sustainability energy development^29,39,40. Using National Land Cover Data (NLCD) created by the Multi-Resolution Land Characteristics (MRLC) consortium, we classify the surface cover within each project footprint (Fig. 5a, b) and report both contemporaneous and historical results. In 2019, 59% of projects in our sample are located on land classified as developed (including low, medium, and high intensity), accounting for 65% of total installed capacity. Using NLCD from 2001, when U.S. solar development remained in its infancy, only 1.6% of total installed capacity in our dataset was located on land classified as developed; however, 38% of projects were developed on land that was classified as cultivated crops. Our results confirm previous findings that rural anthromes (e.g., cropland) and biomes (e.g., grassland) that are close to human population are the most likely to have been previously sited for PV projects^29,41. This outcome points to the potential for optimizing the dual use of land (e.g, agrivoltaics) to mitigate impacts on local ecological systems while sustaining growth of solar PV. Results indicate that PV projects are being built on various types of land, indicating the need to carefully design dual use systems and consider the sustainability of siting decisions across a variety of local environments. Approximately 14% of existing PV projects in our study area now are located on cultivated cropland, which could be retrofitted into dual use systems to obtain techno-ecological synergies²⁹. An additional 15% of the PV projects utilize rooftop modules, which occupy surfaces within the built environment for solar energy. It can be argued that they do not directly transform additional land, highlighting their potential advantage as a land sparing opportunity. Land sparing opportunities exist but are underrepresented in our sample—PV deployment has in many places driven measurable changes in land cover.
These figures show the number and total installed capacity of existing solar PV projects delineated in our dataset across various land cover types (woody wetlands and grassland/herbaceous are new classification types that are not present in the 2001 data). The number atop each bar shows the aggregate installed capacity of all solar PV projects constructed on each specific land cover type. The dark bars indicate developed land cover types while the medium and light bars indicate agricultural and undeveloped land cover types, respectively. a This graph shows the land covers for all PV plant location in 2019. b This graph shows the land covers for the same PV plant location in 2001.
Our systematic methodology enabled the first digital dataset containing project-level land information of solar PV connected to the U.S. Western Interconnection using deep learning. Using this dataset, we contributed a detailed demonstration of land-solar interactions in the region, resulting in novel records of the land use efficiency and land transformation of existing PV facilities with varying geographic and structural characteristics. Our findings confirm that land-solar interactions are highly dependent on local solar output potential and designs of panel mounting systems^42,43. We build upon prior studies that apply deep learning to detect and characterize the area of solar PV as the first to characterize the land required by solar PV projects in their entirety as opposed to the solar arrays alone^29,33,34,36. In doing so, we provide information on solar land use that is rarely seen in existing data—the entire project site of ground-mounted solar PV projects—and present a more complete picture of land-solar interactions.
A novel observation differentiating our findings from prior research is that facilities with higher performance, dual axis tracking mounting systems, are less efficient in terms of project-level land-use when compared to their fixed/single-axis counterparts. We attribute this finding to the much larger clearing space needed to install individual dual-axis solar arrays (e.g., to avoid inter-panel shading), which thereby reduce the panel density⁴⁴. These results point to the need for analyses of land-solar interactions to go beyond the array footprint alone to better understand impacts related to the full project area. Additionally, our data enabled an important classification of the land cover occupied by PV installations, presenting an important approach that highlights the benefits of using the built environment to reduce ecosystem impacts⁴⁵. The majority of the land used for PV projects utilizes land that previously belongs to the rural anthromes and biomes and is now developed, suggesting that solar energy projects are most frequently sited near human population and that land sparing outcomes can be achieved by siting projects to make use of both dual use agrivoltaics and the built environment. Definitive conclusions about the aggregate land-use impacts of solar PV facilities would necessitate ex-ante land cover classification for the same plots of land and potential knock-on effects of local solar energy developments on land use in other areas.
Importantly, our study represents the first attempt to quantify land use metrics using deep learning results, making it the first large-scale analysis of this kind that examines land transformation and land use efficiency. The results demonstrate the potential of using machine learning techniques to perform tasks that are prohibitively difficult and labor intensive (e.g., annotation of PV project site). The combination of GIS and machine learning tools is contributing valuable instruments for consistent collection and analysis of the land-use implications of energy developments. A nascent stream of research has shown the prospects of utilizing open-access map such as the OpenStreetMap to detect and create land-use inventories^46,47. Such approaches have the benefit of coverage of large areas but have so far relied on relatively coarse imagery and remain challenged in achieving the high-precision boundary delineation required for intra-facility solar land-use analysis. Crowdsourced data, such as OpenStreetMap, also may be limited in completeness, geographic coverage, and consistency, and has resulted in incomplete solar inventories⁴⁷. High-resolution aerial imagery such as that we use here more closely represents the ground truth of energy system development, ensuring a replicable approach to producing up-to-date analyses of the rapid development of solar PV. While semantic segmentation for PV detection is increasingly well established, we extend the field by integrating land use efficiency and land transformation metrics^21,48.
While this study primarily focuses on quantifying land-solar interactions for PV projects, results have broad implications for data-driven decision-support in fields such as LCA and energy systems planning. LCAs of electricity generation, including solar PV, are increasingly including spatiotemporal inventories⁴⁹. Site-level data, such as the inventory we produce here, can enable analyses that account for improved geographic representativeness and variability of impacts across projects. While spatial variability is important for solar PV, public inventories have been limited to date^11,33,34,36. Future work could further enhance geospatial information by better characterizing transmission lines, access roads and distribution infrastructure^50,51.
With rapid growth in solar energy, it is vitally important to improve land management and preserve ecological health while mitigating climate change⁴⁵. A large-scale map of existing PV projects can inform a better understanding of land-solar interactions, including other social and geographical aspects of existing land-based economies and ecosystems³⁰. Our methods and results contribute important decision-support for planning, policymaking, and analyzing the deployment of PV across states in the Western Interconnection. By publishing an open-access, georeferenced inventory with a carefully documented methodology, we establish a replicable approach to quantifying land-solar interactions, serving as an important benchmark for understanding the land implications of energy transitions with high growth in solar power.
The study area for this research project is the U.S. part of the Western Interconnection which includes the states of Arizona, California, Colorado, Idaho, Montana, Nevada, New Mexico, Oregon, Utah, Washington, Wyoming, and the El Paso County of Texas. We identified potential solar facility locations in our study area and collect annual generation data using the EIA facility data from 2019 (e.g., EIA-860 and EIA-923), which include electricity generating units with installed capacity of at least 1 megawatt. The land area of each solar PV project is determined using a combination of deep learning technology and aerial imagery analysis, the latter of which involved manual annotation to inform the deep learning model. The aerial imagery used in this study comes from the National Agriculture Imagery Program (NAIP) database administered by the U.S. Department of Agriculture. The NAIP images are acquired at a resolution of <1-meter ground sample distance across the entire US during agricultural growing seasons, representing the more accurate depiction of land-use compared to crowdsourced map databases used in some existing studies⁴⁷. The image database is updated with a 3-year cycle—the one used in this study is published in 2019. The NLCD is a detailed mapping of the US with consistent and relevant land cover information from 2001 to 2019 with a two-to-three-year interval, created by a group of U.S. agencies named the MRLC. Data for the year 2001 and 2019 are used to better elucidate the change in land cover before and after almost all solar PV installations in our study area are built. The pixel resolution for NLCD is 30 m—suitable for solar PV projects which typically occupy around 200,000 square meters of land and much finer than what has been used in previous findings²⁹.
We compare our estimates against two datasets on the land use of solar PV facilities for the conterminous U.S. published by the USGS in 2016 and jointly by the USGS and the Lawrence Berkeley National Laboratory in 2023^30,31. The 2016 USGS dataset identifies locations of solar PV panels in the conterminous US in EIA utility-scale facilities data from 2015 and digitizes the footprint area of each solar array using aerial imagery from NAIP from the same year. In total, they documented that there were 740 U.S. solar PV facilities (1207 arrays) by 2015. Since the scope of our study is limited to the U.S. part of the Western Interconnection, we have a different study area than the two existing datasets.
The solar PV projects examined in our study are distributed across different climate regions with vastly different landscapes. Before deploying a deep learning model for image annotation, which has been shown to be a useful tool for land-use inventory-making and analysis, we need to ensure the performance of the model^{29,33,34,36,37,47}. Our deep learning model needs to be able to capture the distinct features of solar PV systems against the varying backdrops of land cover types. To that end, we begin by training the model using a diverse set of manual annotations. Each solar PV power plant is located using coordinates obtained from the U.S. Energy Information Administration on NAIP. The image of the plant and its immediate surrounding area is then cropped from the aerial imagery and saved for annotation. Specifically, a randomly selected group of 100 projects were manually annotated at first for the development of a deep learning model. The entire portfolio was then automatically annotated by the deep learning model. The product generated by the model then goes through manual post-processing before being compiled into a master shapefile.
Table 1 shows a summary of the six annotation classes we use for solar PV projects. In this solar PV annotation framework, PV panels, inverters, and other infrastructures (e.g., battery storage stations or buildings) are all within the project site. In addition, there will be no overlaps between PV panels and inverters, nor between external access roads and project sites.
In this annotation framework, we only consider (visually distinguishable) external access roads as a separate class, and consider internal roads for operation and maintenance as part of the project site without further classification. In addition, we consider the overall project site as a permanent site clearing and do not distinguish other temporary land use, which usually recovers or is cleared once the project construction is finished, and the project is in operation.
The PV panels class can be annotated in multiple ways based on different decision rules on polygon delineation. Here we adopt a “visually distinguishable segments” rule to delineate solar PV panels—separating clusters of PV panels by evidence of relatively larger empty space that can be potentially utilized for other purposes (e.g., dirt road for maintenance vehicles). We decide on this rule in order to separate PV panels from adjacent support infrastructure such as inverters and acknowledge spaces that are reserved for maintenance access, while also categorizing small intervals between PV panels as part of the PV panels class since these parcels of land would not be used for other purposes. For instance, Fig. 6 shows an example of the annotation based on this “visually distinguishable” rule. Here solar PV panels are separated from on-site inverters.
This is part of the AV Solar Ranch One (240 megawatts) in CA. The color codes can be found in Table 1.
To understand the impacts on PV panel areas using different polygon delineation rules, we compared the land area impacts with four different decision rules. The four rules being tested are based on three delineation methods: 1) the PV panel-only rule, which only accounts for physical PV panels and separates small intervals between panels; 2) the visually distinguishable segments rule, which is the decision rule we adopt as described above; and 3) the quantitative distance rule, which uses a quantitative measure to decide on segmentation, i.e., if panels have a distance larger than the threshold, they are categorized as different polygons, otherwise categorized as a single polygon. For the 3^rd method, we use both 10 m and 30 m as decision rules to test, where the 30-m rule is adopted in the previous USGS study. An example of the annotation for the PV panels class based on different decision rules can be found in Fig. 7.
Sample project: Pine Tree Solar Project (8.5 megawatts), CA.
We compare the PV panel land areas for four solar PV projects using these four different decision rules. Potential differences between the decision rules that can be used in this methodology are illustrated in Table 2. Only accounting for PV panels results in about a 50 percent reduction in land footprint compared to using a 30-m rule in these sample projects. However, other than the PV panel-only method, the land area impacts using the other three decision rules are very similar, suggesting minor deviations when switching between these three rules. According to the example shown here, the difference in land area impacts of using different methods would be greater with larger size of projects.
While developing annotated images for solar PV, we also measure two distances in order to justify decision rules for the paper: minimum distance segmented and maximum distance not segmented. An example of the two distances being measured is shown in Fig. 8. The minimum distance segmented represents the minimum distance where two segments of PV panels are separated (due to inverters, roads, or other visually distinguishable features), whereas the maximum distance not segmented represents the maximum distance between PV panels that are aggregated as a single segment.
Example of minimum distance segmented and maximum distance not segmented.
A deep learning-based network was trained to perform automatic segmentation of the PV panels based on a set of 100 solar PV facilities which is manually annotated according to rules detailed in the section above. We specifically chose U-Net as the deep network for segmentation as it has been seen to work well for segmentation-based applications in aerial imaging where precise boundaries are needed⁵². U-Net is an encoder-decoder architecture where the encoder takes in the image as input and converts it to a lower dimensional latent space. These features are then forwarded to a decoder which converts the low-dimensional features to the segmentation map. The encoder path resembles a traditional convolutional neural network (CNN) architecture, commonly seen in more recent end‑to‑end deep‑learning frameworks for mapping solar arrays^33,34,36. It consists of a series of convolutional and pooling layers that progressively reduce the spatial dimensions of the input image while extracting high-level features. Each stage in the encoder path typically consists of two convolutional layers followed by a max pooling operation. The decoder uses a combination of upsampling (transposed convolution) and concatenation operations to recover the spatial information lost during the encoding process. Each stage in the decoder path consists of an upsampling layer followed by two convolutional layers. U-Net incorporates skip connections between corresponding stages of the encoder and decoder paths. These skip connections allow the network to retain low-level details during the upsampling process, enabling precise localization of objects in the segmented output. Repeating such computation process using our training set of 100 annotations enables the model to take in any image and outline different objects within the image (e.g., PV panels). We develop the model by splitting the data into a 70-15-15 split for the train-validation-test. The model is developed using Python for coding, the Pytorch library for training, and a NVIDIA A100 GPU cluster for computation. We use an Adam optimizer with a learning rate of 0.001 with a batch size of 16. The model is trained for 300 epochs until convergence.
Having developed our deep learning model, we start examining the land-use profile of our entire portfolio of solar PV projects. First, we geo-locate each solar plant in WECC using publicly available information from the EIA and download the corresponding NAIP images using USGS EarthExplorer. In rare cases where the coordinates in EIA surveys do not point to the actual location of the solar PV projects, we try to visually locate said projects using NAIP images or Google Maps. Areas of interest are then determined by creating a buffer area around the perimeter of each plant, with the buffer distance ranging from over 50 meters to include possible land use near the perimeter. The NAIP images within the buffer boundary are split into multiple images, each with a resolution of 1024 pixels by 1024 pixels. Each bundle of images for one plant is then fed into our deep learning model for examination, the outcome of which is merged back into a single shape file delineating the captured land use profile for each plant. Finally, we manually post-process each shape file, correcting any imperfections and collecting data for further land use analysis (Fig. 9).
Example of the workflow of our methodology at the plant-level process.
In the end, our model performs well in terms of identifying image features that resemble empty spacing between PV panels as well as PV panels themselves. However, we do not provide land-use information on ancillary infrastructure in the final results because our model recognizes the various types and shapes of ancillary infrastructure across different landscapes with relatively poorer accuracy than it does panels. To identify any systemic bias in the annotations of panels and project sites, we benchmark our results pre and post manual corrections utilizing the Intersection over Union (IoU) score as our performance metric. Our analysis shows a mean IoU score of 0.822 with high statistical significance (p = 0.0062) and low variance (0.035). The results indicate little difference between the images pre and post manual correction, meaning that there was not much manual correction needed during the post-processing process.
The typical nomenclature used in existing literature quantifies land-use efficiency in terms of land area “transformation” and “occupation” metrics. The former is used to evaluate the installation impact of energy systems, while the latter is used to consider the aggregate impact of system installation and operation^9,11,26. The inherent difficulty in reporting land-use results under the current nomenclature lies in the lack of standardization in calculations of land area transformation and occupation. Currently, there are dozens of different terms used in existing studies to describe essentially the same two metrics, the majority of which are used both in capacity-based calculations and generation-based calculations with varying timescales^19,48. The variation in methods and reporting metrics makes land-use estimations largely incomparable and, in turn, obscures the accurate translation of scientific findings into policy decisions²¹.
For this study, we proposed the usage of “land transformation” which has been one of the most common metrics used in life cycle assessment literature and other fields^{11,53,54,55,56}. Each project was analyzed by calculating the land area altered per unit of energy produced in the assumed 30-year lifetime using the following equation:
The land area of each project was estimated in terms of the total project footprint and the PV panel footprint. The former refers to all the land within the project site boundaries while the latter refers to the land occupied by the solar arrays. We also differentiate land-use efficiency by calculating capacity-based and generation-based land use efficiency metrics, which are needed to provide a holistic review of the theoretical and real-world land-use efficiency of solar PV facilities¹⁴. Importantly, our published data enable users to transparently calculate metrics as appropriate to their analyses⁴⁸. Each project was analyzed by calculating the land area altered per unit of installed nameplate capacity or annual electricity production in 2019 using the following equations:
The post-processed inventories (panel and project footprints) that are used produce the findings of this study are available on Zenodo (https://doi.org/10.5281/zenodo.17058755). High-resolution aerial imagery used for annotation and validation were obtained from the U.S. Department of Agriculture National Agricultural Imagery Program (https://earthexplorer.usgs.gov/). National Land Cover Database land-cover layers were obtained from the Multi-Resolution Land Characteristics consortium (https://www.mrlc.gov/). Project metadata (installed capacity, generation, etc) were compiled from the Form-860 (https://www.eia.gov/electricity/data/eia860/) and Form-923 (https://www.eia.gov/electricity/data/eia923/) by the U.S. Energy Information Administration. The 2016 solar photovoltaic panel annotations used for comparison were obtained from the U.S. Geological Survey (https://doi.org/10.5066/F79S1P57). The 2023 solar photovoltaic footprint data produced by the U.S. Geological Survey and the Lawrence Berkeley National Laboratory were obtained from https://doi.org/10.5066/P9IA3TUS).
The deep-learning model and code to reproduce the figures are archived alongside the data on Zenodo (https://doi.org/10.5281/zenodo.17058756).
Creutzig, F. et al. The underestimated potential of solar energy to mitigate climate change. Nat. Energy 2, 17140 (2017).
Article  Google Scholar 
Jacobson, M. Z. et al. 100% clean and renewable wind, water, and sunlight all-sector energy roadmaps for 139 countries of the world. Joule 1, 108–121 (2017).
Article  Google Scholar 
Luderer, G. et al. Assessment of wind and solar power in global low-carbon energy scenarios: an introduction. Energy Econ. 64, 542–551 (2017).
Article  Google Scholar 
Dale, S. (analysis). BP Statistical Review of World Energy 2021. BP p.l.c. (2021). Available at: https://www.bp.com/content/dam/bp/business-sites/en/global/corporate/pdfs/energy-economics/statistical-review/bp-stats-review-2021-full-report.pdf?.
Haegel, N. M. et al. Terawatt-scale photovoltaics: transform global energy. Science 364, 836–838 (2019).
Article  CAS  Google Scholar 
Hernandez, R. R. et al. Environmental impacts of utility-scale solar energy. Renew. Sustain. Energy Rev. 29, 766–779 (2014).
Article  Google Scholar 
MacKay, D. J. C. Solar energy in the context of energy use, energy transportation and energy storage. Philos. Trans. R. Soc. Math. Phys. Eng. Sci. 371, 20110431 (2013).
Google Scholar 
Smil, V. Power Density Primer: Understanding the Spatial Dimension of the Unfolding Transition to Renewable Electricity Generation (Part I – Definitions) (2010). Available at: https://moodle2.units.it/pluginfile.php/553999/mod_folder/content/0/smil-article-power-density-primer.pdf?.
Turney, D. & Fthenakis, V. Environmental impacts from the installation and operation of large-scale solar power plants. Renew. Sustain. Energy Rev. 15, 3261–3270 (2011).
Article  Google Scholar 
Jordaan, S. M., Lee, J., McClung, M. R. & Moran, M. D. Quantifying the ecosystem services values of electricity generation in the US Chihuahuan Desert: a life cycle perspective. J. Ind. Ecol. 25, 1089–1101 (2021).
Article  Google Scholar 
Fthenakis, V. & Kim, H. C. Land use and electricity generation: a life-cycle analysis. Renew. Sustain. Energy Rev. 13, 1465–1474 (2009).
Article  Google Scholar 
Bukhary, S., Ahmad, S. & Batista, J. Analyzing land and water requirements for solar deployment in the Southwestern United States. Renew. Sustain. Energy Rev. 82, 3288–3305 (2018).
Article  Google Scholar 
Capellán-Pérez, I., de Castro, C. & Arto, I. Assessing vulnerabilities and limits in the transition to renewable energies: land requirements under 100% solar energy scenarios. Renew. Sustain. Energy Rev. 77, 760–782 (2017).
Article  Google Scholar 
Hernandez, R. R., Hoffacker, M. K. & Field, C. B. Land-use efficiency of big solar. Environ. Sci. Technol. 48, 1315–1323 (2014).
Article  CAS  Google Scholar 
Majumdar, D. & Pasqualetti, M. J. Analysis of land availability for utility-scale power plants and assessment of solar photovoltaic development in the state of Arizona, USA. Renew. Energy 134, 1213–1231 (2019).
Article  Google Scholar 
Mauro, G. & Lughi, V. Mapping land use impact of photovoltaic farms via crowdsourcing in the Province of Lecce (Southeastern Italy). Sol. Energy 155, 434–444 (2017).
Article  Google Scholar 
Wu, X. et al. Unveiling land footprint of solar power: a pilot solar tower project in China. J. Environ. Manag. 280, 111741 (2021).
Article  Google Scholar 
Hernandez, R. R., Jordaan, S. M., Kaldunski, B. & Kumar, N. Aligning climate change and sustainable development goals with an innovation systems roadmap for renewable power. Front. Sustain. 1, 583090 (2020).
Article  Google Scholar 
Hernandez, R. R., Cagle, A. E., Grodsky, S. M., Exley, G. & Jordaan, S. M. Comments on: Land use for United States power generation: a critical review of existing metrics with suggestions for going forward (Renewable and Sustainable Energy Reviews 2021; 143: 110911). Renew. Sustain. Energy Rev. 166, 112526 (2022).
Article  Google Scholar 
Wachs, E. & Engel, B. Land use for United States power generation: a critical review of existing metrics with suggestions for going forward. Renew. Sustain. Energy Rev. 143, 110911 (2021).
Article  Google Scholar 
Cagle, A. E. et al. Standardized metrics to quantify solar energy-land relationships: a global systematic review. Front. Sustain. 3, 1035705 (2023).
Article  Google Scholar 
Fritsche, U. et al. Energy and Land Use – Global Land Outlook Working Paper (UNCCD, 2017). https://doi.org/10.13140/RG.2.2.24905.44648.
IINAS. Selected Results from GEMIS 4.95: Electricity Generation (IINAS, 2017).
Trainor, A. M., McDonald, R. I. & Fargione, J. Energy sprawl is the largest driver of land use change in United States. PLOS ONE 11, e0162269 (2016).
Article  Google Scholar 
Groesbeck, J. G. & Pearce, J. M. Coal with carbon capture and sequestration is not as land use efficient as solar photovoltaic technology for climate neutral electricity production. Sci. Rep. 8, 13476 (2018).
Article  Google Scholar 
Jordaan, S. M. The land use footprint of energy extraction in Alberta. PhD thesis, University of Calgary (2010) https://doi.org/10.11575/PRISM/3520.
Ong, S., Campbell, C., Denholm, P., Margolis, R. & Heath, G. Land-Use Requirements for Solar Power Plants in the United States. NREL/TP-6A20-56290, 1086349, https://doi.org/10.2172/1086349 (2013).
Jordaan, S. M. et al. Understanding the life cycle surface land requirements of natural gas-fired electricity. Nat. Energy 2, 804–812 (2017).
Article  Google Scholar 
Kruitwagen, L. et al. A global inventory of photovoltaic solar energy generating units. Nature 598, 604–610 (2021).
Article  CAS  Google Scholar 
Carr, N. B., Fancher, T., Freeman, A. T. & Battles Manley, H. Surface area of solar arrays in the conterminous United States. U.S. Geological Survey data release (2016). Available at: https://doi.org/10.5066/F79S1P57.
Fujita K. S. et al. United States Large-Scale Solar Photovoltaic Database (ver. 2.0, August 2024). U.S. Geological Survey https://doi.org/10.5066/P9IA3TUS (2024).
Bolinger, M. & Bolinger, G. Land requirements for utility-scale PV: an empirical update on power and energy density. IEEE J. Photovolt. 12, 589–594 (2022).
Article  Google Scholar 
Yu, J., Wang, Z., Majumdar, A. & Rajagopal, R. DeepSolar: a machine learning framework to efficiently construct a solar deployment database in the United States. Joule 2, 2605–2617 (2018).
Article  Google Scholar 
Wang, Z., Arlt, M.-L., Zanocco, C., Majumdar, A. & Rajagopal, R. DeepSolar++: understanding residential solar adoption trajectories with computer vision and technology diffusion models. Joule 6, 2611–2625 (2022).
Article  Google Scholar 
Malof, J. M., Bradbury, K., Collins, L. M. & Newell, R. G. Automatic detection of solar photovoltaic arrays in high resolution aerial imagery. Appl. Energy 183, 229–240 (2016).
Article  Google Scholar 
Hou, X. et al. SolarNet: a deep learning framework to map solar plants in China from satellite imagery. In ICLR 2020 Workshop Tackling Clim. Change Mach. Learn (ICLR, 2020).
Camilo, J., Wang, R., Collins, L. M., Bradbury, K. & Malof, J. M. Application of a semantic segmentation convolutional neural network for accurate automatic detection and mapping of solar photovoltaic arrays in aerial imagery. Preprint at https://doi.org/10.48550/ARXIV.1801.04018 (2018).
IEA. Renewables 2019. https://www.iea.org/reports/renewables-2019 (2019).
Hoffacker, M. K., Allen, M. F. & Hernandez, R. R. Land-sparing opportunities for solar energy development in agricultural landscapes: a case study of the Great Central Valley, CA, United States. Environ. Sci. Technol. 51, 14472–14482 (2017).
Article  CAS  Google Scholar 
Cagle, A. E. et al. The land sparing, water surface use efficiency, and water surface transformation of floating photovoltaic solar energy installations. Sustainability 12, 8154 (2020).
Article  CAS  Google Scholar 
Maguire, K., Tanner, S. J., Winikoff, J. B., Williams, R., & United States. Department of Agriculture. Economic Research Service. Utility-Scale Solar and Wind Development in Rural Areas: Land Cover Change (2009-20). https://handle.nal.usda.gov/10113/8374829https://doi.org/10.32747/2024.8374829.ers (2024).
Leccisi, E., Raugei, M. & Fthenakis, V. The energy and environmental performance of ground-mounted photovoltaic systems—a timely update. Energies 9, 622 (2016).
Article  Google Scholar 
Kurnik, J., Jankovec, M., Brecl, K. & Topic, M. Outdoor testing of PV module temperature and performance under different mounting and operational conditions. Sol. Energy Mater. Sol. Cells 95, 373–376 (2011).
Article  CAS  Google Scholar 
Jensen, A. R., Sifnaios, I., Furbo, S. & Dragsted, J. Self-shading of two-axis tracking solar collectors: impact of field layout, latitude, and aperture shape. Sol. Energy 236, 215–224 (2022).
Article  Google Scholar 
Hernandez, R. R. et al. Techno–ecological synergies of solar energy for global sustainability. Nat. Sustain. 2, 560–568 (2019).
Article  Google Scholar 
Stowell, D. et al. A harmonised, high-coverage, open dataset of solar photovoltaic installations in the UK. Sci. Data 7, 394 (2020).
Article  Google Scholar 
Dunnett, S., Sorichetta, A., Taylor, G. & Eigenbrod, F. Harmonised global datasets of wind and solar farm locations and power. Sci. Data 7, 130 (2020).
Article  Google Scholar 
Turkovska, O. et al. Methodological and reporting inconsistencies in land-use requirements misguide future renewable energy planning. One Earth 7, 1741–1759 (2024).
Article  Google Scholar 
Jordaan, S. M., Combs, C. & Guenther, E. Life cycle assessment of electricity generation: a systematic review of spatiotemporal methods. Adv. Appl. Energy 3, 100058 (2021).
Article  Google Scholar 
Lovering, J., Swain, M., Blomqvist, L. & Hernandez, R. R. Land-use intensity of electricity production and tomorrow’s energy landscape. PLOS ONE 17, e0270155 (2022).
Article  CAS  Google Scholar 
Daniela-Abigail, H.-L. et al. Life cycle assessment of photovoltaic panels including transportation and two end-of-life scenarios: Shaping a sustainable future for renewable energy. Renew. Energy Focus 51, 100649 (2024).
Article  Google Scholar 
Ronneberger, O., Fischer, P. & Brox, T. U-Net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (eds, Navab, N., Hornegger, J., Wells, W. M. & Frangi, A. F.) 9351 234–241 (Springer International Publishing, 2015).
Dai, T. et al. Land resources for wind energy development requires regionalized characterizations. Environ. Sci. Technol. 58, 5014–5023 (2024).
Article  CAS  Google Scholar 
Koellner, T. & Scholz, R. W. Assessment of land use impacts on the natural environment: Part 2: generic characterization factors for local species diversity in Central Europe. Int. J. Life Cycle Assess. 13, 32–48 (2008).
Google Scholar 
Koellner, T. et al. UNEP-SETAC guideline on global land use impact assessment on biodiversity and ecosystem services in LCA. Int. J. Life Cycle Assess. 18, 1188–1202 (2013).
Article  Google Scholar 
De Baan, L., Alkemade, R. & Koellner, T. Land use impacts on biodiversity in LCA: a global approach. Int. J. Life Cycle Assess. 18, 1216–1230 (2013).
Article  Google Scholar 
Download references
This research was supported by the Alfred P. Sloan Foundation (grant number: G-2023-19646). We acknowledge the feedback on our research and generosity of time provided by an external expert review panel, comprised of government, nonprofit, industry, and academics. Experts included Timothy Skone, Jim Kuiper, Garvin Heath, Matthew Bailey, Barry Woertz, Benjamin Riggan, Rama Chappella, James O’Sullivan, Christopher Newman, Tim Hayes, Jane Long, Armond Cohen, and Anders Johnson. We also acknowledge Johns Hopkins support from IDIES (Gerard Lemson and Alexander Szalay) and in GIS training (Bonni Wittstadt). Tao Dai was a postdoctoral scholar working on this project and provided helpful assistance to the lead author based on his concurrent work. Views and opinions presented in this article are those of the authors.
Department of Agricultural and Applied Economics, University of Wisconsin-Madison, Madison, WI, USA
Siyuan Hu
Department of Environmental Health and Engineering, Johns Hopkins University, Baltimore, MD, USA
Yinong Sun
Department of Land, Air and Water Resources, University of California Davis, Davis, CA, USA
Rebecca R. Hernandez
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, USA
Jeya Maria Jose Valanarasu & Vishal M. Patel
Department of Civil Engineering and the Trottier Institute for Sustainability in Engineering and Design, McGill University, Montreal, QC, Canada
Sarah M. Jordaan
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
S.H.: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Data curation, Resources, Writing—original draft, Writing—review & editing, Visualization. Y.S.: Methodology, Writing—original draft. R.R. H.: Methodology, Visualization, Writing—review and editing. J.M.J.V.: Methodology, Software, Validation, Formal analysis, Writing—original draft. V. M. P.: Methodology, Software, Supervision. S. M. J.: Conceptualization, Methodology, Formal analysis, Data curation, Resources, Writing—original draft, Writing—review & editing, Visualization, Supervision, Project administration, Funding acquisition. All authors contributed to the work and have approved the final version of the manuscript.
Correspondence to Sarah M. Jordaan.
The authors declare no competing interests.
Communications Earth & Environment thanks and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Martina Grecequet. A peer review file is available.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Hu, S., Sun, Y., Hernandez, R.R. et al. Quantifying land-use metrics for solar photovoltaic projects in the western United States. Commun Earth Environ 6, 1006 (2025). https://doi.org/10.1038/s43247-025-02862-5
Download citation
Received: 10 April 2025
Accepted: 02 October 2025
Published: 15 December 2025
Version of record: 15 December 2025
DOI: https://doi.org/10.1038/s43247-025-02862-5
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Communications Earth & Environment (Commun Earth Environ)
ISSN 2662-4435 (online)
© 2025 Springer Nature Limited
Sign up for the Nature Briefing: Anthropocene newsletter — what matters in anthropocene research, free to your inbox weekly.

source