Narrow urban rivers remain challenging for Sentinel-2-based chlorophyll-a (Chl-a) retrieval because mixed pixels, adjacency effects, and bank shadows can contaminate spectral signals, while in situ observations are often limited. This study develops an integrated workflow that integrates a high-confidence urban-river water mask with explicit shadow suppression to constrain feature extraction and sample selection, cross-resolution pseudolabel transfer by mapping a GF-1 Chl-a field to the Sentinel-2 10 m grid and stratifying it to form pseudolabeled samples for augmentation, and a leakage-controlled out-of-fold stacking regressor with an XGBoost metalearner to improve stability under data scarcity. On the in situ test set (n = 6), the GF-1 retrieval achieves R-2 = 0.86 and RMSE = 3.32 mu g/L. Using mapped pseudolabels for Sentinel-2 training, the model attains R-2 = 0.84 and RMSE = 3.88 mu g/L on a held-out pseudolabel set (n = 40), while the in situ only baseline yields R-2 = 0.70 and RMSE = 4.47 mu g/L on the separate in situ test set. These results indicate that cross-resolution pseudolabel augmentation and leakage-controlled ensembling can be effective under limited field supervision; the pseudolabel evaluation reflects internal consistency rather than independent ground truth.

Rainfed rice is critical for global food security yet remains highly vulnerable to climate change. Traditional linear models often fail to decouple the complex, non-linear interactions between drivers and crop growth dynamics. This study integrates multi-source satellite observations with an interpretable machine learning framework to quantify the relative contributions and nonlinear response patterns of environmental factors associated with rainfed rice growth across South and Southeast Asia. Water availability and vapor pressure deficit (VPD) were the primary factors associated with rainfed rice growth in Bangladesh, Cambodia, and India, whereas temperature and VPD were the dominant factors in Myanmar and Thailand. Specifically, elevated VPD exerted a consistent inhibitory effect on growth, while solar radiation acted as a consistent promoter. Crucially, cumulative precipitation and precipitation frequency demonstrated distinct threshold effects, shifting from beneficial to inhibitory beyond specific inflection points. Future projections based on the CMIP6 dataset indicate that future rice growth will tend to decline, with a markedly stronger decline under SSP5-8.5 compared to SSP2-4.5. These findings clarify the environmental thresholds governing rainfed rice growth, providing a scientific basis for region-specific adaptation strategies under a warming climate.

The temporal mismatch between structural (canopy green-up) and functional (physiological carbon uptake) spring phenology in forests remains poorly explored. This study quantified this temporal mismatch (Delta SOS) across Northern Hemisphere (NH) forests during 2001-2021 by integrating multi-source datasets with an interpretable machine-learning framework. We further developed a bidirectional Gated Recurrent Unit (GRU) model to project future changes in Delta SOS under different scenarios. Results showed that although both spring phenological events advanced significantly, the magnitude and direction of Delta SOS varied strongly among forest types, with structural phenology lagging behind functional phenology in evergreen forests but preceding it in deciduous forests. Pixellevel trend analysis showed that Delta SOS changed significantly in 59.5 % of forested areas, with significant decreases prevailing in evergreen needleleaf forests (ENF, 55.6 %), evergreen broadleaf forests (EBF, 69.9 %), and deciduous needleleaf forests (DNF, 63.3 %) but increases dominating deciduous broadleaf forests (DBF, 66.1 %). The spatiotemporal changes in this mismatch mainly stemmed from divergent sensitivities of functional and structural phenology to the same environmental drivers, reflected by driver-specific differences in response magnitude and direction, with temperature and solar radiation emerging as the dominant factors. Future projections indicated that significant Delta SOS trends were projected across 63.2 % (SSP2-4.5) and 73.8 % (SSP5-8.5) of forests, with absolute rates of change under SSP5-8.5 being approximately two to three times higher than those under SSP2-4.5. Unlike the historical pattern, DNF was dominated by increases under both scenarios (75.0 % and 72.5 %). These findings suggest that failing to account for this mismatch may bias future terrestrial carbon sink estimates.

Accurate maize yield prediction is crucial for food security and risk management under climate change. Existing methods often fail to capture agricultural management heterogeneity, model nonlinear multi-factor interactions, and dynamically characterize factor contributions across growth stages, limiting prediction accuracy, robustness, and practical relevance. To address these limitations, we propose an interpretable machine learning model that incorporates nitrogen and phosphorus fertilizer application, agricultural mechanization level, and the adoption rate of genetically modified varieties to represent management-related yield heterogeneity. In addition, interactions among multiple yield drivers are explicitly modeled, and changes in factor contributions are analyzed across growth stages. The model was evaluated using a 22-year dataset from 2001 to 2022 across 1,069 counties in the United States Corn Belt. Results showed that the model outperformed representative benchmark models in prediction accuracy while exhibiting greater robustness under extreme weather conditions. It achieved a coefficient of determination (R2) of 0.81 and a root mean square error (RMSE) of 0.99 tonnes per hectare (t ha−1), improving R2 by at least 0.05 and reducing RMSE by at least 0.13 t ha−1. Dynamic interpretability analysis further revealed stage-specific associations between yield predictions and key environmental factors, with patterns consistent with excessive early-season wetness and water–heat stress during the flowering and grain-filling stages. These findings may inform agricultural management practices. Overall, this study offers an interpretable and robust model for regional maize yield prediction under climate change.

Accurate assessment of bifacial photovoltaic (PV) potential is essential for global energy planning. However, optimizing installation parameters remains challenging due to the complex geometry of ground-reflected radiation. This study addresses this issue by presenting a framework for mapping global bifacial PV potential, which involves determining the optimal tilt and azimuth angles. We integrate a height-corrected view factor model into a hybrid optimization algorithm that combines particle swarm optimization with gradient-based local search. This overcomes the limitations of traditional cross-strings methods by explicitly accounting for module mounting height. We simulate the optimal configuration and energy yields of bifacial PV worldwide using multi-source global datasets, including CERES radiation, ERA5 meteorological reanalysis and MODIS albedo products. The simulation is validated against in-situ irradiance and energy yield measurements from bifacial experimental sites (NREL and Sandia). Our model outperforms standard models (pvlib, PVRT, and RT) in irradiance simulation. For the energy conversion step, we incorporate a bifacial thermal model and a mismatch loss factor, which further improves the accuracy. The optimized installation angles reveal distinct geographic patterns driven by latitude and ground albedo. The optimal tilt angles for bifacial systems are generally 0-10 degrees lower than those for monofacial counterparts, particularly in high-latitude and high-albedo regions. Furthermore, our performance analysis shows that optimized bifacial systems achieve a global average energy gain of 10.92% (ranging from 0% to 46.17%) compared to monofacial systems. Angular optimization can provide an additional yield increase of up to 10% in high-latitude regions. This work reveals the global patterns in bifacial PV performance, providing valuable insights for PV system design, resource assessment, and energy yield forecasting.

Conventional assessments of carbon-pool change at photovoltaic (PV) facilities typically assign static carbon-density values to broad land-cover transitions. They do not explicitly represent microclimatic modifications associated with PV arrays. This limitation may be important in drylands, where reflectance-related thermal conditions and vegetation responses vary among land-cover settings and precipitation conditions. Carbon-pool changes within 189 confirmed PV facilities covering 254 km2 in Inner Mongolia are evaluated using a conventional static land-cover carbon-density approach M1 and a microclimate-adjusted approach M2. M2 incorporates a signed microclimate adjustment term δmicro into M1 to represent net responses associated with PV-related differences in reflectance, thermal conditions, and vegetation-related conditions. Across land-cover transitions, δmicro ranged from +36.2% for bare-land-to-grassland transitions to -12.6% for grassland-to-bare-land transitions. Field-based evaluation at 42 PV facilities showed that M2 had higher agreement with measured carbon densities than M1 (R2 = 0.730 vs. 0.624; RMSE = 2.130 vs. 2.512 kg C m-2; MAE = 1.706 vs. 1.896 kg C m-2). M2 also showed lower site-level absolute errors in paired tests (p = 0.041 for the paired t-test and p = 0.037 for the Wilcoxon signed-rank test). Relative microclimate adjustments are largest in the low-precipitation group, whereas the central tendency of estimated carbon-pool change is greater in the high-precipitation group. These findings indicate that microclimate-adjusted accounting can improve carbon-pool assessment at PV facilities and provide context-sensitive information for post-construction vegetation management and land-use screening in dryland grasslands.

Accurate mapping of stony desert is important for dryland monitoring but remains limited by spectral confusion and the high computational cost of conventional remote sensing methods. This study presents a global mapping framework based on Geospatial Foundation Models (GFMs), combined with a Few-Shot Similarity Retrieval (FSR) approach that operates in the learned embedding space. By leveraging pretrained GFM representations, the proposed framework enables similarity-based retrieval in latent feature space and reduces reliance on large labeled datasets required by supervised learning approaches. The FSR with GFM approach achieves approximately 90% overall accuracy using only 750 reference samples, approaching the observed performance of SDL models trained with large labeled datasets. In terms of deployment efficiency, GFM-based FSR substantially reduces global mapping time compared with the supervised pipeline under the experimental configuration used in this study. Cross-regional experiments further showed that FSR could identify SD surfaces across geographically distant arid regions using reference samples from a single region. A 10 m resolution global stony desert map for 2025 is produced using this framework, with an estimated global SD of 6.34 ± 0.15 million km2. These results highlight the effectiveness of foundation model representations for large-scale geomorphological mapping and demonstrate a practical solution for arid land surface extraction under limited annotation conditions, with potential applications in energy infrastructure planning and soil inorganic carbon assessment.

Climate change is intensifying global energy demands and amplifying exposure to extreme heat. Building fa & ccedil;ade-integrated photovoltaics (FIPV) present a largely untapped opportunity to supply renewable electricity while enhancing urban climate resilience. Here we show that deployable FIPV systems worldwide could generate 732.5 +/- 4.5 TWh of electricity annually, based on a global synthesis of building datasets, climate projections and fa & ccedil;ade-scale simulations, with theoretical bounds of 8.9-7,671.3 TWh under conservative-to-optimistic assumptions. Although FIPV deployment costs exceed those of conventional photovoltaics, over 80% of urban districts exhibit lifetime expenditure savings due to combined electricity generation and cooling-load reductions. Under a gradual S-curve adoption reaching upper-bound potential by 2050, FIPV could deliver cumulative emission reductions of up to 37.7 GtCO2, corresponding to 0.0519 +/- 0.0111 degrees C of avoided warming under currently announced national policies. These results identify FIPV as a complementary mitigation-adaptation strategy, highlighting the need for targeted policies to address regional and economic disparities in climate-resilient urban transition.

Solar photovoltaic (PV) systems play a crucial role in addressing the growing demand for clean energy and mitigating climate change impacts. However, PV system performance is heavily influenced by the incident solar radiation on panel surfaces, with suboptimal tilt angles leading to significant power losses. Despite the critical importance of tilt angle optimization, many existing PV installations worldwide operate suboptimally due to simplified estimation methods or lack of site-specific optimization. This study presents a novel hybrid approach combining empirical and computational methods to determine optimal annual and monthly PV panel tilt angles using long-term hourly ERA5 reanalysis radiation data. Our results validate the effectiveness of ERA5 data for global tilt angle optimization, demonstrating a strong correlation with established cubic relations. Analysis of spatial and temporal patterns of optimized tilt angles reveals the influence of latitude, local atmospheric conditions, and seasonal variations on optimal PV panel inclination. A comprehensive assessment of the global PV inventory in 2018 shows that 44.6% of installed capacity is located in regions with solar power losses exceeding 1%, resulting in a total loss of 6154 GWh yr-1-equivalent to Luxembourg's annual electricity consumption. Comparison between optimized tilt angles and those estimated using empirical cubic schemes reveals significant discrepancies in some regions, with annual power losses surpassing 3% when using empirical methods. These findings underscore the importance of accurate, location-specific tilt angle optimization to minimize solar power losses and maximize global PV inventory performance. Our research highlights the potential for substantial energy yield improvements through widespread adoption of optimized tilt angles in PV system design and retrofitting, contributing to enhanced renewable energy production and accelerated progress towards global sustainability goals.

Accelerating energy transition towards renewables is central to net-zero emissions. However, building a global power system dominated by solar and wind energy presents immense challenges. Here, we demonstrate the potential of a globally interconnected solar-wind system to meet future electricity demands. We estimate that such a system could generate ~3.1 times the projected 2050 global electricity demand. By optimizing solar-wind deployment, storage capacity, and trans-regional transmission, the solar-wind penetration could be achieved using only 29.4% of the highest potential, with a 15.6% reduction in initial investment compared to a strategy without interconnection. Global interconnection improves energy efficiency, mitigates the variability of renewable energy, promotes energy availability, and eases the economic burden of decarbonization. Importantly, this interconnected system shows remarkable resilience to climate extremes, generation outages, transmission disruptions, and geopolitical conflicts. Our findings underscore the potential of global interconnection in enabling high renewable penetration and guiding sustainable energy transitions.

Compound drought and heatwave events (CDHWs) are increasing in frequency across dryland ecosystems, yet their impacts on vegetation remain insufficiently studied. In this study, we develop a new CDHWs index that integrates a four-day standardized soil moisture index with heatwaves index. Employing an interpretable machine learning model, we quantitatively analyzed the effects of various environmental factors on vegetation during CDHWs. Results demonstrate significant upward trends in CDHW frequency, intensity, and maximum duration across global drylands. Analysis indicates that 65% of CDHWs negatively affect vegetation, 20% produce positive effects, and 15% result in minimal impact. Accumulated local effects and partial dependence plot analyses identify maximum temperature as the primary factor driving negative vegetation responses, while soil moisture emerges as the predominant factor associated with positive responses. Projections under multiple coupled model intercomparison project phase 6 scenarios (SSP245 and SSP585) indicate continued increases in CDHW frequency, intensity, and maximum duration through 2054. These findings underscore the necessity for fine-scale monitoring of compound extremes and offer actionable insights for adaptive management in dryland regions.

Many vegetation phenological models predominantly rely on temperature, overlooking the critical roles of water availability and soil characteristics. This limitation significantly impacts the accuracy of phenological projections, particularly in water-limited ecosystems. We proposed a new approach incorporating soil enthalpy-a comprehensive metric integrating soil moisture, temperature, and texture-to improve phenological modeling. Using an extensive dataset combining FLUXNET observations, solar-induced fluorescence (SIF), and meteorological data across the Northern Hemisphere (NH), we analyzed the relationship between soil enthalpy and vegetation phenology from 2001 to 2020. Our analysis revealed significant temporal trends in soil enthalpy that corresponded with changes in leaf onset date (LOD) and leaf senescence date (LSD). We developed and validated a new soil enthalpy-based model with optimized parameters. The soil enthalpy-based model showed particularly strong performance in autumn phenology, improving LSD simulation accuracy by at least 15% across all vegetation types. For shrub and grassland ecosystems, LOD projections improved by more than 12% compared to the temperature-based model. Future scenario analysis using CMIP6 data (2020-2054) revealed that the temperature-based model consistently projects earlier LOD and later LSD compared to the soil enthalpy-based model, suggesting potential overestimation of growing season length in previous studies. This study establishes soil enthalpy as a valuable metric for phenological modeling and highlights the importance of incorporating both water availability and soil characteristics for more accurate predictions of vegetation phenology under changing climatic conditions.

Inaccurate edge detection is a common challenge in the segmentation of household rooftop photovoltaic (PV) systems from remote sensing images, which hinders the accurate retrieval of PV distribution information critical for planning and managing PV development. A widely adopted solution is to incorporate an additional edge detection task into a joint-task learning framework to enhance edge perception. However, existing joint-task learning methods often struggle to accurately detect PV edges and lack effective mechanisms for distinguishing PV edges from those of similar objects. To address the above challenges, we develop a novel joint-task learning framework. This framework introduces a Scale Adaptive Module (SAM) that dynamically adjusts the receptive field of edge features based on the PV actual size and shape, enabling precise detection of PV edges with varying shapes and sizes. In addition, a Position Guidance Module (PGM) is proposed based on the intrinsic relationship between the PV segmentation task and the edge detection task. The PGM not only guides the edge detection task to focus on identifying the semantic edges of PVs using the distribution information from the segmentation task but also enhances the ability of the segmentation task to accurately locate PVs in complex backgrounds by utilizing the backward gradient from the edge detection task. Multiple rounds of repeated experiments on the Duke and IGN datasets demonstrate the framework's superior performance. Compared to other models, the proposed framework significantly improves the detection accuracy of various PV edges, achieving the best performance in household rooftop PV segmentation with an Intersection over Union (IoU) of 77.4 %. This study provides valuable insights into the accurate acquisition of household rooftop PV information and offers a promising solution for object segmentation tasks facing the challenge of inaccurate edge extraction.

Accurate detection of warming trends is crucial for both mitigation and adaptation strategies. While satellite observations provide high spatial resolution temperature data, cloud contamination creates gaps that require filling methods-a process that can bias warming trend calculations. Traditional gap-filling approaches, though accurate for absolute temperatures, consistently underestimate warming trends. This study presents a new data assimilation method that integrates moderate-resolution imaging spectroradiometer derived air temperatures with ERA5-Land reanalysis data to achieve unbiased warming trend detection while maintaining high spatial resolution. We validate our method against weather station data on the Tibetan Plateau (TP) and compare it with four mainstream gap-filling approaches: 1) temporal, 2) spatial, 3) spatio-temporal, and 4) multisource fusion-based methods. Our results show that while all methods, including ours, perform similarly in terms of absolute temperature accuracy (with root-mean-square error around 1.66 degrees C and R around 0.99), a critical difference emerges in warming trend estimation. Traditional gap-filling methods show negative biases in warming trends, but our assimilation-based approach almost completely eliminates these biases when validated against both station-level data and elevation-binned averages. The integrated air temperature for the TP reveals significant warming patterns, particularly in glacier regions, although the overall warming rate (0.022 degrees C/yr) is lower than that indicated by station data alone (0.026 degrees C/yr). This difference likely reflects the ability of our method to capture warming trends across the diverse terrain of the plateau, not just at station locations. This improvement in trend estimation, combined with the method's ability to maintain high spatial resolution, represents a significant advance in the use of satellite-derived data for climate change analysis.

Accurate projections of future surface solar radiation (SSR) are important for assessing the impacts of climate change and the potential of solar energy. However, climate models from the Coupled Model Intercomparison Project Phase 6 (CMIP6) exhibit notable uncertainties in SSR projections. This study aims to develop a high quality monthly SSR dataset during 1850–2100 by synthesizing CMIP6 model projections and satellite-derived retrievals using a Bayesian Linear Regression (BLR) method. Five CMIP6 models are selected based on their historical performance in simulating SSR. The BLR method assigns gridded weights to each model based on how well the historical simulations matched the satellite-based SSR product (called ISCCP‒ITP‒CNN) over the period 1983–2014. The weighted multi-model ensemble is calculated to generate a synthesized long-term SSR dataset. Evaluation against ground-based observations during historical periods (1960–2017) shows that the synthesized SSR outperforms individual CMIP6 models and their original multi-model mean, with a reduced RMSE from 32 to 36 W/m2 to 25 W/m2 and a bias from 5 to 13 W/m2 to −1 W/m2 on monthly scales. The spatial patterns also agree well with the ISCCP‒ITP‒CNN (1983–2018). The high-resolution (0.1° × 0.1°) synthesized SSR dataset provides monthly projections over historical experiments and four future shared socio-economic pathway (SSP) scenarios (SSP126, SSP245, SSP370, and SSP585) during 1850–2100, representing future SSR changes and associated climate impacts. The dataset is expected to enhance simulations of land surface processes and solar energy applications under a variety of future climate scenarios.

AbstractPhotovoltaic (PV) installations are a leading technology for generating green electricity and reducing carbon emissions. Roofing highways with solar panels offers a new opportunity for PV development, but its potential of global deployment and associated socio‐economic impacts have not been investigated. Here, we combine solar PV output modeling with the global highway distribution and levelized cost of electricity to estimate the potential and economic feasibility of deploying highway PV systems worldwide. We also quantify its co‐benefits of reducing CO2 equivalent emissions and traffic losses (road traffic deaths and socio‐economic burdens). Our analysis reveals a potential for generating 17.58 PWh yr−1 of electricity, of which nearly 56% can be realized at a cost below US$100 MWh−1. Achieving the full highway PV potential could offset 28.78% (28.21%–29.1%) of the global total carbon emissions in 2018, prevent approximately 0.15 million road traffic deaths, and reduce US$0.43 ± 0.16 trillion socio‐economic burdens per year. Highway PV projects could bring a net return of about US$14.42 ± 4.04 trillion over a 25‐year lifetime. To exploit the full potential of highway PV, countries with various income levels must strengthen cooperation and balance the multiple socio‐economic co‐benefits.

The impacts of various drought types on autumn phenology have yet to be extensively explored. We address the influence of pre-season agricultural and meteorological droughts on autumn phenology in the Northern Hemisphere. To this end, enhanced autumn phenology models incorporating drought factors was developed, contributing to a deeper understanding of these complex interactions. The study reveals that there was no significant trend of advancement or delay in the End of Season (EOS) across the Northern Hemisphere based on SIF estimates from 2001 to 2020. The cumulative and delayed impacts of pre-season agricultural drought on EOS were found to be more pronounced than those associated with meteorological drought. The analysis of various evaluation indexes shows that the performance of the Cooling Degree Days (CDD) model incorporating the Standardized Soil Moisture Drought Index (CDDSSMI) in simulating EOS in the Northern Hemisphere is >14 % higher than that of the standard CDD model. Additionally, the performance of the CDD model with the Standardized Precipitation Index (CDDSPI) in simulating EOS in the Northern Hemisphere is improved by >5.6 % compared to the standard CDD model. A comparison of future EOS projections across various models reveals that the CDD model significantly overestimates EOS in different scenarios (SSP245 and SSP585). The CDDSSMI model projects EOS approximately 7 days earlier than the CDD model, and the CDDSPI model projects EOS approximately 5 days earlier than the CDD model. This study highlights the diverse impacts of drought types on plant autumn phenology and underscores the significance of parameterizing drought impacts in autumn phenology models.

The optimal tilt angle for photovoltaic (PV) systems is crucial for maximizing solar energy capture. China's diverse climate and geography pose challenges for tilt angle optimization. This study addresses the challenges by using a data-driven approach to determine grid-specific optimal tilt angles across China. Long-term ERA5 hourly solar radiation data and an optimization procedure are used to calculate the annual and monthly tilt angle that maximizes the total solar radiation received over different time periods (up to ten years) for every grid in China. Sensitivity analyses indicate that the optimized tilt angle varies by up to 10° depending on the time period considered, highlighting the importance of long-term radiation data. It is shown that four latitude-dependent schemes in China cannot accurately match the optimized tilt angles, emphasizing the need for optimization through the use of local solar radiation. The difference between our optimized tilt angles and ones via a best-performing latitude scheme makes for an estimated PV power loss of approximately 1.11 TWh/year based on China's PV installations in 2018, equivalent to the PV power generation of Philippines or Portugal in the same year. The presented data-driven framework contributes to China’s PV industry by determination of optimal fixed tilt angles.

The precise location and size of distributed photovoltaics (PVs) is critical to infer the actual installed capacity and assess the remaining PV generation potential, and is therefore the cornerstone of strategic planning for distributed PV deployment. However, identifying small-scale distributed PVs in complex contexts from high spatial resolution remote sensing (HSRRS) images to obtain their information remains an issue. In this study, we propose an advanced deep learning model, called PV Identifier, to enhance the identification accuracy of small-scale PV systems from HSRRS images. PV Identifier uses a fine-grained feature layer (FFL) compatible with the size of PVs to improve the detection capability of the small-scale distributed PVs. At the same time, it effectively distinguishes between PVs and similar background using a novel semantic constraint module (SCM). We test PV Identifier on a distributed PV dataset in California. Experiments show that the inclusion of the FFL positively affects the model's sensitivity to small distributed PVs. Specifically, the PV Identifier with the FFL increases the Recall of identifying residential rooftop PVs by 1.9% compared to the model without the FFL. In addition, the integration of the SCM effectively improves the model's ability to locate residential rooftop PVs in complex environments, resulting in a 1.8% increase in the corresponding Precision. Compared to the four commonly used segmentation models, PV Identifier exhibits superior identification performance for residential rooftop PVs and commercial and industrial PVs, with an Intersection over Union (IoU) of 74.1% and 89.3%, respectively, which is at least 4.1% and 1.8% higher than other models. Overall, PV Identifier provides a viable solution to the problem of identifying small-scale distributed PV in complex backgrounds from HSRRS images.

Accurate information on the location, shape, and size of photovoltaic (PV) arrays is essential for optimal power system planning and energy system development. In this study, we explore the potential of deep convolutional neural networks (DCNNs) for extracting PV arrays from high spatial resolution remote sensing (HSRRS) images. While previous research has mainly focused on the application of DCNNs, little attention has been paid to investigating the influence of different DCNN structures on the accuracy of PV array extraction. To address this gap, we compare the performance of seven popular DCNNs—AlexNet, VGG16, ResNet50, ResNeXt50, Xception, DenseNet121, and EfficientNetB6—based on a PV array dataset containing 2072 images of 1024 × 1024 size. We evaluate their intersection over union (IoU) values and highlight four DCNNs (EfficientNetB6, Xception, ResNeXt50, and VGG16) that consistently achieve IoU values above 94%. Furthermore, through analyzing the difference in the structure and features of these four DCNNs, we identify structural factors that contribute to the extraction of low-level spatial features (LFs) and high-level semantic features (HFs) of PV arrays. We find that the first feature extraction block without downsampling enhances the LFs’ extraction capability of the DCNNs, resulting in an increase in IoU values of approximately 0.25%. In addition, the use of separable convolution and attention mechanisms plays a crucial role in improving the HFs’ extraction, resulting in a 0.7% and 0.4% increase in IoU values, respectively. Overall, our study provides valuable insights into the impact of DCNN structures on the extraction of PV arrays from HSRRS images. These findings have significant implications for the selection of appropriate DCNNs and the design of robust DCNNs tailored for the accurate and efficient extraction of PV arrays.