Soil fertility in mowing-grazing grasslands of the Qinghai–Tibet Plateau is jointly influenced by hydrothermal conditions, terrain, grassland type and management disturbance, yet reliable identification of relative low-fertility risk remains difficult under sparse sampling. We developed a soil fertility index (SFI) from 1096 topsoil samples collected at 0–30 cm during 2023–2025 and integrated static ecological background, conventional climate, multi-year seasonal hydrothermal statistics, soil hydraulic attributes and self-supervised temporal embeddings. Generalisation was evaluated using spatial-block, leave-year-out and leave-ecological-zone-out validation, while area of applicability (AOA) and spatial-block conformal prediction constrained spatial interpretation. Multi-year seasonal hydrothermal statistics were the most stable predictor source, increasing spatial-block R² by 0.2435 over the static-background model. Strict pre-sampling tests showed that a 10-year seasonal window best balanced explained variance, prediction error and rank consistency. The final deployable model achieved R² = 0.5632, RMSE = 0.1105, Spearman = 0.4473 and AUC = 0.696 under spatial-block validation. GRU/Transformer embeddings provided only supplementary information and did not replace interpretable hydrothermal statistics. Grid evaluation consistently identified Zone 2 as the main concentration of relative low-SFI risk. AOA_Mech, AOA_Full734 and AOA_Core50 covered 82.9%, 81.9% and 85.9% of grid cells, respectively. Normalized conformal prediction increased 90% interval coverage to 0.926, with wider intervals. Natural-background residuals showed stronger negative deviations in Zone 1, whereas human activity and grazing intensity provided only correlative clues. This framework supports spatial prioritisation for alpine grassland fertility monitoring, sample densification and field-based management verification.

Accurate estimation of regional cultivated alfalfa yield is critical for forage management and precision agriculture. However, traditional empirical inversions are frequently limited by spectral saturation at high biomass densities and the spatiotemporal trade-offs of single-sensor platforms. To address these challenges, this study proposes a structural simulation-driven deep learning (SSDL) framework. The methodology integrates multi-source optical satellite imagery from Sentinel-2, Landsat 8, and the Moderate Resolution Imaging Spectroradiometer (MODIS) with intermediate morpho-structural traits simulated by the Decision Support System for Agrotechnology Transfer (DSSAT) crop model. The cross-validation results demonstrate that the SSDL architecture achieved robust predictive accuracy (R2 = 0.67, RMSE = 0.74 t/ha), mitigating the substantial overfitting and generalization gaps observed in baseline machine learning models (R2 = 0.58, RMSE = 0.88 t/ha). A feature importance analysis indicates that sentinel-2 data represented the largest contribution to the predictive model, accounting for 33.9 % of the overall feature weighting due to its spatial resolution and red-edge sensitivity. By leveraging fused inputs, the framework dynamically prioritizes structurally insensitive red-edge and biochemical proxies. This prioritization effectively captures the biochemical accumulation governing peak biomass, bypassing the asymptotic saturation of broadband structural indices. The incorporation of simulated traits serves as a physical constraint, anchoring the data-driven spectral correlations and enabling high-resolution mapping of intra-field yield heterogeneity. While the framework notably improves the estimation of macroscopic canopy architecture, the analysis also acknowledges the persistent uncertainties of utilizing two-dimensional passive optical data to resolve three-dimensional vertical plant height and sub-surface biomass partitioning. Overall, the SSDL framework provides a scalable and biophysically constrained approach for continuous regional forage monitoring.

Maize leaf diseases in field environments often exhibit large variations in lesion scale, irregular morphology, blurred boundaries, and complex backgrounds. These factors pose challenges for existing detection models, particularly in detecting small lesions and achieving precise bounding-box localization. To address these issues, this study proposes LSL-YOLO11n, a maize leaf disease detection model based on the YOLO11n framework. The proposed model improves feature representation, localization quality modeling, and bounding-box regression to enhance disease detection performance under complex field conditions. Experiments were conducted on a dataset containing 15,119 images and 29,366 annotated instances across eight categories, including seven maize disease categories and healthy leaves. To evaluate the effectiveness of the proposed model, ablation experiments, comparative experiments with mainstream object detection models, and visual detection analyses were carried out. The ablation results show that the improved components contribute positively to the overall detection performance. LSL-YOLO11n achieves a Precision of 84.4%, Recall of 73.9%, and mean Average Precision (mAP) of 83.3%, which is 3.1 percentage points higher than that of the baseline YOLO11n model. Compared with YOLOv8n, YOLOv9t, YOLOv10n, and YOLOv12n, the proposed model improves mAP by 4.7, 3.3, 5.3, and 10.9 percentage points, respectively. The visual detection results further indicate that LSL-YOLO11n performs more stably in complex backgrounds and small-lesion scenarios. These findings provide technical support for rapid maize disease recognition and intelligent field monitoring.

Soil fertility in mown-grazed grasslands on the Qinghai–Tibet Plateau reflects hydrothermal conditions, terrain, grassland type, and management disturbance. We developed a soil fertility index (SFI) from 1037 topsoil samples collected at 0–30 cm during 2023–2025. The modelling framework combined static ecological background variables, conventional climate indicators, multi-year seasonal hydrothermal statistics, soil hydraulic attributes, and self-supervised temporal embeddings. Spatial-block, year-held-out, and ecological-zone-held-out validation were used with area-of-applicability (AOA) analysis and spatial-block conformal prediction. Multi-year seasonal hydrothermal statistics supplied the strongest predictive information and increased spatial-block R2 by 0.2435 relative to the static-background model. A 10-year seasonal window gave the best empirical balance among explained variance, prediction error and rank consistency. The final deployment model achieved R2 = 0.5632, RMSE = 0.1105, Spearman = 0.4473 and AUC = 0.696 under spatial-block validation. These values support regional screening and sampling prioritisation, not local deterministic diagnosis or site-level management prescriptions. Grid prediction identified Zone 2 as the main concentration of relative low-SFI risk. Full-sample AOA schemes covered 82.9%, 81.9% and 85.9% of the system-evaluation grid cells, whereas deployment-grid AOA coverage was lower under the stricter deployment setting. This contrast separates feature-space support from operational grid support. Split conformal prediction achieved 90% coverage close to the nominal level (PICP = 0.899, MPIW = 0.264). Natural-background residuals separated relative low-SFI risk from local deviations below expected natural conditions. The framework provides a reproducible screening tool for regional prioritisation and follow-up field verification in alpine grasslands.

Accurate and timely monitoring of alfalfa crude protein (CP) using multi-source remote sensing data is essential for optimizing livestock management and improving agricultural productivity. While multispectral satellite imagery provides extensive spatial coverage, its coarse spectral resolution lacks the characteristic absorption bands associated with CP, limiting its ability to support precise CP estimation. In contrast, UAV-based hyperspectral data offer rich spectral information but remain costly and difficult to scale for large-area monitoring. Existing multi-source fusion methods struggle to effectively reconcile the spatial and spectral constraints of different sensors due to inherent data limitations, operational complexity, or the absence of physical constraints. To address these challenges, this study integrates physically based radiative transfer models (RTMs) into a deep learning architecture to develop a physics-guided spectral super-resolution (SSR) framework capable of reconstructing high-fidelity hyperspectral data from Sentinel-2 (S2) multispectral data, thereby enhancing the accuracy of alfalfa CP predictions using S2 imagery. The results demonstrate that the RTMs-SSR module significantly improves spectral reconstruction fidelity, achieving optimal performance (RMSE of 0.91 and MAE of 0.70). Notably, the optimal CP prediction was achieved by combining the RTMs-guided SSR with LASSO feature selection (R2 = 0.87, RMSE = 1.62%), outperforming models based on original S2 data (R2 = 0.40, RMSE = 3.08%). SHAP interpretability analysis further validated the framework by identifying the SWIR and NIR as the primary drivers of accurate reconstruction. These results demonstrate the significant potential of RTMs-guided deep learning for SSR, offering a cost-effective and scalable alternative to UAV-based hyperspectral data acquisition for agricultural monitoring, enabling high-precision crop quality monitoring across extensive agricultural landscapes.

Alfalfa (Medicago sativa) is an important forage source for grassland agricultural development; developing accurate and efficient methods for alfalfa identification and yield estimation using remote sensing is of considerable interest. However, the traditional methods of identifying large areas of crops and yield estimation have some problems, such as the limited spatial resolution of remote sensing data and a strong dependence on training data. In this study, using Sentinel-2 high-resolution imagery and the Google Earth Engine (GEE) platform, we constructed a cloud-free normalized difference vegetation index (NDVI) time-series dataset and proposed an effective method for alfalfa feature extraction and yield estimation. The results show that: (1) the producer’s accuracy, user’s accuracy, overall accuracy, and Kappa coefficient of alfalfa identification using the trough recognition algorithm were 98.51%, 91.67%, 94.26%, and 0.88, respectively. The total area of cultivated alfalfa identified in the study area in 2020 was estimated at 46,793.21 hm2, and was mainly distributed in the northern region of the Qilian Mountains. (2) NDVI showed a highly significant correlation with alfalfa hay yield, and the power function regression model performed best, with an R2 greater than 0.65. (3) The annual unit hay yield of four alfalfa cuttings was estimated at 17,497.55–32,962.10 kg/hm2, with a total hay yield of 4.838 × 108 kg and an average hay yield of 4464.95 kg/hm2. The proposed method has significant application potential for automated and rapid remote sensing-based identification and yield estimation of large-scale alfalfa cultivation.

Accurate and timely estimation of alfalfa (Medicago sativa L.) yield across diverse field conditions is crucial for optimizing forage management and ensuring agricultural productivity. While hyperspectral remote sensing offers rich spectral information for non-destructive monitoring, translating this high-dimensional data into robust yield predictions remains challenging. Purely data-driven machine learning (ML) models often function as black boxes that tend to overfit to localized environmental noise. Conversely, mechanistic models, despite their biological robustness, face difficulties in directly assimilating optical data across varied agricultural landscapes. To address these limitations, this study proposes a coupled biophysics-guided deep learning (BGDL) framework that synergizes the mechanistic principles of the DSSAT-CSM model with hyperspectral observations. The framework employs a dual-stream architecture, integrating a mentor model pre-trained on simulated physical data (DSSAT-PROSAIL) with an observer model trained on empirical field data encompassing diverse irrigation and harvest scenarios. Analytical evaluations confirm that the BGDL framework effectively mitigates the severe overfitting observed in traditional ML models, achieving robust generalization for yield prediction (R2 = 0.66, RMSE = 0.90 t/ha). By leveraging the mentor stream to enforce biological consistency, the framework addresses the non-unique inversion problem typical of physical reflectance algorithms, facilitating reliable parameter retrieval on synthetic data (R² = 0.95). SHAP interpretability analysis further validated the framework, confirming that physical guidance directs the model to prioritize biologically causal spectral features—specifically chlorophyll absorption (630 nm) and canopy water content (1141 nm). Furthermore, the framework exhibited notable data efficiency, maintaining reliable predictive performance even when empirical training data was reduced to 60%. Overall, these results demonstrate the significant potential of biophysics-guided deep learning to accurately capture complex interactions between canopy spectra and crop physiology, providing a scientifically and data-efficient approach for precision crop monitoring across complex agro-ecosystems.

Precise monitoring of alfalfa nitrogen content is essential for optimizing forage quality and harvest timing. While hyperspectral remote sensing offers a non-destructive alternative to traditional field sampling, conventional empirical inversion models often fail to account for the confounding influence of three-dimensional canopy morpho-structural dynamics, which frequently leads to substantial overfitting. Although explicitly incorporating structural parameters can stabilize these inversions, acquiring in situ morphological data remains operationally challenging for large-scale applications. To address this limitation, this study proposes the Morphology-Aware Latent Embedding Network (MALENet), a hybrid framework that integrates physiological and morphological mechanisms with hyperspectral data. The framework utilizes the Decision Support System for Agrotechnology Transfer (DSSAT) coupled with the PROSAIL radiative transfer model to generate a structurally rich synthetic hyperspectral prior. A customized one-dimensional convolutional neural network, augmented with a Squeeze-and-Excitation (SE) attention mechanism, is pre-trained on this synthetic data to distill complex ecophysiological knowledge into a dense latent representation. This mechanistically derived embedding is subsequently fused with empirical hyperspectral measurements collected across diverse field conditions. Cross-validation indicated that the latent space effectively reconstructed key morphological drivers, most notably Leaf Area Index (R² = 0.98) and leaf weight (R² = 0.97). By isolating the biochemical nitrogen signal from structural noise, MALENet exhibited enhanced predictive performance compared to conventional machine learning and standard deep learning models, achieving strong predictive accuracy and stability (Test R² = 0.71, RMSE = 0.30). Ultimately, this integrated approach mitigates structural confounding, offering a robust and scalable methodology for monitoring the spatiotemporal dynamics of alfalfa quality in precision agriculture.

The nutrient status of forage during the senescent stage in an alpine grassland ecosystem of the Tibetan Plateau provides important information for the precise supplementary feeding of grazing livestock. During the senescent stage, the application of remote sensing technology to understand the spatial pattern of forage quality parameters accurately is crucial for grazing management. However, few studies have used relatively high-resolution multispectral images to map forage nutritional value, palatability, and digestibility during the senescent stage; the potential and feasibility of this approach have not been effectively confirmed. The present study aims to map the forage crude protein (CP), acid detergent fiber (ADF), neutral detergent fiber (NDF), crude fiber (CF), nitrogen-free extract (NFE), crude fat (EE), relative feed value (RFV) and organic matter digestibility (OMD) in alpine grassland during the senescent stage based on Sentinel-2 spectral variables and auxiliary data (i.e. topography, meteorology and soil). The results indicate that the spectral bands favorable for forage quality parameter retrievals are generally distributed in the red and shortwave infrared regions; the validation accuracy of the proposed optimal model varied among quality parameters (V-R2 of 0.28-0.69), with the model performing well for CP (V-R2 = 0.62, V-RMSE = 1.21%), EE (V-R2 = 0.46, V-RMSE = 0.35%), CF (V-R2 = 0.40, V-RMSE = 2.92%), and NFE (V-R2 = 0.69, V-RMSE = 3.60%) and unsatisfactorily (V-R2 of 0.28-0.39) for NDF, ADF, RFV, and OMD. Moreover, some vegetation indices derived from the wavebands of red, red-edge and near infrared regions and environmental factors such as elevation, radiation and temperature contributed significantly to forage quality parameter estimations. Overall, our study demonstrates that the integration of environmental variables and multispectral remote sensing data can not only improve the prediction performance of forage quality parameters, but also enable large-scale mapping of these parameters, with potential applications for the development of high-precision monitoring methods for forage nutritional status during the senescent stage of alpine grassland.

Assessing the state of the grassland-livestock balance is of great practical significance for continuously optimizing grazing management and ensuring the sustainable use of alpine grasslands. However, obtaining fine, dynamic, and continuous information of the grassland-livestock balance in alpine grassland is greatly hindered due to limitations in precision, timeliness, and spatial scale of two key indicators, theoretical livestock carrying capacity (TLCC) and actual livestock carrying capacity (ALCC). On the one hand, this study constructed grassland aboveground biomass (AGB) estimation models using the multispectral remote aerial vehicle (RAV) data and Sentinel-2 data in the northeast of Tibetan Plateau based on the nonnegative matrix factorization (NMF)-based differentiated fusion method. The annual high-precision TLCC spatial distribution dataset was obtained through simulation from 2019 to 2023. On the other hand, this study constructed livestock density estimation models based on the livestock statistical data of the smallest administrative unit and various gridded environmental and socioeconomic data from 2019 to 2022; it thus achieved the fine spatialization of the regional-scale ALCC. The spatiotemporal dynamics of TLCC and ALCC sequences were comprehensively analyzed, and the grassland-livestock balance status and its driving factors from 2019 to 2022 were further evaluated in the study area. We concluded that: 1) the R-2 of the optimal RAV-based and Sentinel-2-based grassland AGB models reached 0.88 and 0.77, with corresponding root mean square error (RMSE) of 552.39 and 692.41 kg/ha, respectively; 2) differentiated fusion with RAV spectral data significantly enhanced the estimation accuracy of Sentinel-2 data for regional-scale grassland AGB ( R-2 improved from 0.77 to 0.80); 3) the RF-based livestock density model had the best spatialization effect on ALCC ( R-2=0.80 and RMSE =77.34 sheep/km(2)); 4) from 2019 to 2022, the overall status of the grassland-livestock balance was good, with over 65% of the grassland area consistently nonoverloaded or balanced, while severely overloaded grassland area remained below 16.14%; large interannual fluctuations were observed in nonoverloaded and severely overloaded grasslands; the overloaded areas mainly concentrated in the core pasture and the periphery of high population density areas; and 5) livestock density, temperature, and distance to river were the main driving factors of livestock-grassland balance dynamics. Innovatively, an annual 10-m spatial resolution grassland-livestock balance dataset was developed.

Accurate and timely information describing the spatial distribution of alfalfa grasslands is essential for food security and the sustainable development of grass-livestock husbandry. However, the accuracy of remote sensing-based mapping of alfalfa grasslands is typically limited by a lack of ground observation samples and the low spatiotemporal resolution of widely used satellite remote sensing data. In this study, we propose a knowledge-based method to extract high-quality alfalfa training data using integrated multi-source time series satellite remote sensing data. Using the Google Earth Engine platform and machine learning classifiers, we conducted research on mapping the spatial distribution and cutting intensity of alfalfa grasslands in the Ningxia region. The results are as follows: (1) the random forest classification model was constructed by combining automatically generated knowledge-based training sample data, measured sample data, and Sentinel-1, Sentinel-2, and Landsat 8 satellite data, in which achieved good classification accuracy, with an F1 score of 0.97, and an overall accuracy of 0.97; (2) the top three variables with the highest importance are all NIRv (near-infrared reflectance of terrestrial vegetation); (3) spring is the best time to identify alfalfa using multi-source remote sensing data, and (4) through the proposed novel mowing intensity framework for remote sensing mapping of alfalfa cutting intensity here are pronounced differences in cutting intensity in different regions. Overall, this study's results indicate that the proposed method has important application potential for accurate mapping of alfalfa grassland spatial distribution and cutting intensity, as well as alfalfa yield estimation.

Timely and accurate crop mapping is crucial for providing essential data support for agricultural production management. Reliable ground truth samples form the foundation for crop mapping using remote sensing imagery, a task that presents significant challenges in regions with limited sample availability. To address this issue, this study evaluates instance-based transfer learning methods, using the Hexi Corridor as a case study to explore crop mapping strategies in areas with scarce samples. High-confidence pixels from the United States Cropland Data Layer (CDL), along with high-density time series data derived from Sentinel-1, Sentinel-2, and Landsat-8 satellite imagery, as well as key vegetation indices, were selected as training samples for the source domain. Various algorithms, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and TrAdaBoost, were employed to transfer knowledge from the source domain to the target domain for crop type mapping. The results demonstrated that during the transfer learning process using only source domain data—without utilizing any target domain data—the overall classification accuracy reached 73.88%, with optimal accuracies for maize and alfalfa at 88.97% and 85.23%, respectively. As target domain data were gradually incorporated, the total accuracy for all models ranged from 0.77 to 0.92, with F1-scores ranging from 0.76 to 0.92, showing a consistent improvement in model performance. This study highlights the feasibility of employing transfer learning for crop mapping in the Hexi Corridor, demonstrating its potential to reduce labeling costs for target domain samples and providing a valuable reference for crop mapping in regions with limited sample availability.

Accurate and timely assessment of alfalfa nutritional parameters is crucial for optimizing harvest management, maximizing yield, and ensuring high-quality forage in China's Hexi Corridor, a key alfalfa-growing region. UAVbased hyperspectral remote sensing offers a nondestructive and efficient method for monitoring these parameters, providing high-resolution data and covering large areas efficiently. Previous studies have faced challenges related to the scarcity and imbalance of hyperspectral samples and the effective selection of spectral bands for evaluating crop nutrients. Additionally, the simultaneous evaluation of multiple nutrient parameters using a common set of spectral bands has rarely been reported. Least Absolute Shrinkage and Selection Operator (LASSO) is an important method for hyperspectral band selection, but its linear fitting process is challenged by the complex relationship between spectral reflectance and plant properties. In this study, we propose a new band selection strategy that identifies the most informative spectral bands and improves model performance by combining the strengths of both LASSO selection of bands and machine learning's ability to fit complex relationships. To address the issue of imbalanced field samples, we generated high-quality synthetic data using the synthetic minority oversampling technique for regression with Gaussian noise (SMOGN) algorithm. Three machine learning models (ANN, RF, and SVM) were then employed to predict alfalfa nutritional parameters. Our findings show that the proposed synergistic band selection strategy significantly improves model performance, yielding a 14-25 % reduction in RMSE while requiring only 37-59 % of the original spectral bands. By integrating this band selection strategy with the SMOGN method, our optimal model for estimating alfalfa nutrient parameters achieved R2 values of 0.92-0.95 and PRMSE values of 5.1-7.1 %. We observed the importance of the spectral regions around 730 nm and 960 nm for predicting alfalfa quality parameters. This finding suggests that existing satellite platforms such as Sentinel-2 and Landsat could improve the accuracy and efficiency of alfalfa quality monitoring by incorporating these specific spectral bands. Overall, our approach provides a robust and transferable framework for improving the accuracy and reliability of remote sensing-based crop quality monitoring, which is important for optimizing the spectral band configurations of future satellite sensors for precision agriculture.

Grassland, as an essential terrestrial ecosystem, plays a vital role in maintaining ecosystem stability, enhancing carbon sequestration, and promoting biodiversity conservation. Accurate estimation of grassland aboveground biomass (AGB) is essential for effective ecosystem management and sustainable development, especially in environmentally sensitive areas like the Qinghai-Xizang Plateau. This study integrated field-measured grassland AGB data (2003-2023) with MODIS remote sensing data and meteorological variables. A comprehensive benchmarking of 25 machine learning (ML) algorithms was conducted to evaluate their performance. The optimal model was identified through an exhaustive comparison and hyperparameter optimization process, and then applied to estimate grassland AGB across the plateau and analyze its spatiotemporal dynamics. The results showed that Ranger performed best (R2 = 0.6541, RMSE = 512.25 kg DW/ha, MAE = 375.91 kg DW/ha), with OSAVI and precipitation-related variables as the key driving factors. Between 2003 and 2023, the spatial distribution of grassland AGB on the Qinghai-Xizang Plateau exhibited significant heterogeneity, with an overall decreasing trend from the southeast to the northwest. From the perspective of interannual variation, 51.41% of the total area remained stable. With respect to persistent change, 64.71% of the area showed uncertain trends, while 35.29% of the area was characterized by persistent trends. The results of this study provide critical data support and a robust scientific foundation for the ecological protection and sustainable management of grassland resources on the plateau.

Accurate alfalfa yield estimation is essential for efficient forage management and feed planning. We compiled a three-year (2021-2023) dataset of 61 fields, constructing matched Sentinel-2 (S2) and Sentinel-1 (S1) time-series features for each cutting stage (greening, branching, budding, flowering). We benchmarked seven models (long short-term memory (LSTM) networks, Convolutional LSTM (ConvLSTM), convolutional neural networks (CNN), Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Least Absolute Shrinkage and Selection Operator (LASSO)) under two remote sensing image analysis methods (object-based image analysis (OBIA) and pixel-based image analysis (PBIA)) and three input data combinations (S1+S2, S2, and S1). Our results consistently show that OBIA out-performs PBIA for field-scale yield estimation. The S1+S2 fusion also demonstrated superior performance across all metrics compared to single-sensor inputs, highlighting the complementarity of optical and radar data. Additionally, deep learning outperformed machine learning, with LSTM best overall. An annual evaluation using the optimal settings (OBIA, S1+S2, LSTM) revealed small total yield errors (3.97 %-5.90 %), indicating robust generalization across seasonal and inter-annual variations. integrated gradients analysis identified Sentinel-2 SWIR (B11, B12) and Sentinel-1 VV/VH as the most influential variables, with the flowering stage contributing the most. The object-level "sum" statistic improved stability by reflecting with-in-field yield accumulation. Overall, this work presents an accurate, interpretable, and costeffective framework for mapping total alfalfa yield.

China's grasslands, with abundant resources and high carbon sequestration capacity, play an important role in the terrestrial carbon cycle. The root-to-shoot ratio (R/S) reflects the aboveground and belowground carbon allocation patterns of vegetation and is an important parameter for estimating grassland carbon stocks. Previous studies have focused primarily on fixed R/S values from statistical analyses, leading to large uncertainties in areas with high spatial heterogeneity. In this study, a high-accuracy R/S model was constructed using the AutoGluon framework and traditional machine learning (ML) algorithms with 1,367 R/S samples of grassland in China, integrating climate, soil, terrain and spectral features. The results indicated the following: (1) The AutoGluon model achieved superior performance (R-2 = 0.93, RMSE = 18.12) compared to three traditional ML models. (2) Among the 17 natural grassland types, median R/S values ranged from 0.36 to 11.11 and were significantly lower than the corresponding means (0.44 to 13.38), indicating a right-skewed distribution. (3) In terms of climatic zones, the R/S values followed the trend: alpine grasslands > temperate grasslands > warm and tropical grasslands. This study has achieved spatial mapping of natural grassland R/S, offering valuable insights into carbon allocation, ecosystem functioning, and the sustainable management of grassland resources.

Grasslands in China are rich in resources and have a strong carbon sequestration capacity, thus being important in the carbon cycle of terrestrial ecosystems. The root-to-shoot ratio (R/S) reflects the aboveground and belowground carbon allocation patterns of vegetation and is an important parameter for estimating carbon stocks in grasslands. However, researchers looking at grassland R/S have relied primarily on statistical analyses conducted in local areas with limited samples, and there is a lack of relevant studies on the R/S of grassland prediction and spatialization. In this study, a high-accuracy R/S model was constructed based on the AutoGluon framework and traditional machine learning (ML) algorithms using 1,367 R/S of grassland samples in China, combined with climate, soil, terrain and spectral features. Then, the R/S of natural grasslands was simulated at a 1 km resolution, and the R/S spatial distribution characteristics of 18 types of natural grasslands in China were analyzed. The results showed that 1) the estimation accuracy of the R/S model based on the AutoGluon framework was significantly better than that of the other three traditional ML models (R2 = 0.93, RMSE = 18.12). 2) The optimal model consists of 12 variables that are sensitive to changes in the R/S of grasslands, including topographic, climate, spectral, and soil features. 3) Among the 18 types of natural grasslands, alpine steppe had the largest R/S value of 346.52, while warm-temperate tussock had the smallest R/S value of 0.11. For each grassland, the mean R/S values ranged from 0.42 to 14.20, and the median R/S values fell within the range of 0.36 to 11.85. In general, the R/S spatial distribution map of natural grasslands developed in this study is expected to improve the accuracy of net primary productivity and carbon stock estimations in grassland ecosystems.

Alfalfa, a high-quality forage, has good palatability and nutritional value. Neutral detergent fiber (NDF) and acid detergent fiber (ADF) are both key indicators of alfalfa quality. However, the uncertainties in existing studies regarding the sensitive bands and inversion mechanism for NDF and ADF contents estimations have limited the application of high-precision remote sensing-based inversion. In this study, using hyperspectral and Sentinel-2 (S2) multispectral data of cultivated alfalfa in the Hexi Corridor region from 2020 to 2022, we analyze the characteristic spectral band and vegetation indices (VIs) required to estimate the NDF and ADF contents of alfalfa. The key conclusions are as follows. (1) The sensitive bands selected using ASD hyperspectral data are mainly in the blue, green, red-edge, and short-wave infrared (SWIR) regions, while the sensitive bands based on S2 data cover a broader range between the blue and SWIR regions. (2) Among the 21 NDF and 21 ADF models based on ASD data in this study, the optimal models are both artificial neural network (ANN) models constructed by VIs (R2 of 0.80 for both, RMSEs of 2.27% and 1.75% and mean absolute errors (MAEs) of 1.77% and 1.38% for NDF and ADF, respectively). For the S2 data, the optimal models are also ANN-based and constructed using VIs (with R2 values of 0.66 and 0.72, RMSEs of 3.06% and 2.24%, and MAEs of 2.50% and 1.79% for NDF and ADF, respectively. (3) The inversion results using the optimal model indicate that the proportion of alfalfa area in the typical study area with NDF and ADF contents characterized by a supreme grade is greater than 60%. Overall, both the ASD hyperspectral and S2 multispectral data can accurately predict alfalfa NDF and ADF contents. This approach provides an effective technical means by which the management of local alfalfa production may be guided.