Obtaining disaster information promptly post-earthquake provides a powerful scientific basis for deployment of emergency response operation. This study used recorded earthquake data from China to evaluate the usability of minimum-distance similarity models for rapid earthquake impact assessment and extends the approach's application to the assessment of multi-dimensional disaster indicators. Results demonstrate that different similarity models and case quantities yield optimal results for different earthquake impact assessments: Manhattan distance II (3-4 cases) for casualties, standardized Euclidean distance (1-3 or 5-8 cases) for direct economic losses, and Euclidean distance (8-10 cases) for stricken areas.

Maize, as one of the most widely cultivated crops worldwide, is crucial for food security and livelihoods. Accurate dynamic maize mapping is essential for production forecasting and preharvest decision-making. However, current approaches remain limited by incomplete seasonal representation and poor model transferability across growth stages, highlighting the need for an automated and dynamic maize mapping method throughout the growing season. In this study, we first explored the spectral bands that maximally differentiate maize from other crops in terms of water content, chlorophyll levels, and canopy leaf structure during the growing season and proposed a novel normalized difference composite index (NDCI). Based on this index, an automated and dynamic maize identification method that does not rely on crop labels was constructed using a multitemporal Gaussian Mixture Model (GMM), referred to as NDCI-mGMM. The framework was evaluated in five representative maize-growing regions across China, the United States, and France, and further validated in 2021 at two United States sites (Iowa and Georgia) to assess its temporal transferability. Across these regions and years, NDCI-mGMM achieved overall accuracies of approximately 85%, demonstrating stable performance under varying phenological and environmental conditions. Compared with commonly used vegetation indices (Datt99, REP, LSWI, and CIgreen), the NDCI achieved higher F1-scores (by 4-49%) and improved maize separability. Moreover, NDCI-mGMM enabled early-season maize mapping during the tasseling stage-up to two months before harvest-with satisfactory accuracy (F1 >= 79%). As the method operates independently of crop labels, it provides a scalable and temporally transferable framework for in-season maize mapping and early-season area estimation in data-limited regions, thereby supporting timely crop monitoring for production forecasting and preharvest decision-making.

High-resolution (HR) crop maps are the foundation data for conducting precise agricultural monitoring and land use change analysis. However, HR crop mapping remains hindered by costly manual annotations, label noise from low resolution (LR) products, convolutional neural network (CNN) architectural limitations, and spatialspectral constraints of single source remote sensing imagery. LR crop classification products serve as costeffective weak supervision but introduce geometric and semantic label noise. Conventional CNNs struggle with preserving fine spatial details and modeling long-range contexts, while data-level multi-source fusion blurs parcel boundaries. To address these issues, we propose a multisource weakly supervised crop classification framework (MSWCF) that utilizes LR products to guide HR mapping. The MSWCF features dual parallel branches: the HR branch extracts parcel boundaries, while the LR branch captures semantic information. Each branch integrates spatial resolution preserving and Vision Transformer modules to concurrently extract local details and global contexts. A cross-attention module (CAM) adaptively fuses these features, while a noise label removal module suppresses boundary and semantic errors in LR labels. Experiments were conducted in Heilongjiang Province, China. Results show that MSWCF achieves superior performance (Overall Accuracy = 93.8 %) surpassing state-of-the-art baselines. Cross-spatiotemporal application indicates that it has generalization capabilities. Ablation studies confirm that the dual-branch structure and CAM fusion effectively utilize spatial-spectral information, preventing boundary smoothing. Feature visualization verifies that SRP focuses on boundaries while ViT captures global objects. MSWCF enables accurate, fine-grained crop mapping without manual annotations, validating its practicality for largescale HR crop mapping.

Accurate and timely estimation of corn yields in the U.S. (the world's leading corn producer) is crucial for commodity trading and global food security. Most researchers have explored the use of multi-source data, including remote sensing data and climate data, in combination with machine learning models such as Least Absolute Shrinkage and Selection Operator, Support Vector Regression, Random Forest, and Long Short-Term Memory Network (LSTM) for corn yield estimation. In recent years, there have been studies including phenology information to define growth phases (GPs) as the time scale to calculate model variables rater than aggregate time-series data on a fixed time scale to obtain the variables. Although some progress has been made in this yield estimation method, its effectiveness remains to be further comprehensively studied. This study defines GPs to calculate phenology, remote sensing, and climate variables from 2008 to 2022, which were combined with machine learning models to estimate county-level corn yields in the U.S. The results show that: (1) RMSE of all models built by the different combinations of the phenology, remote sensing, and climate parameters range from 740.10 to 1051.31 kg/ha, and these models can achieve ideal yield prediction about two months before harvest; (2) The accuracy of the models constructed with the variables at the GP time scale (predicted R2 values: 0.78-0.83) outperforms that of the models built with the variables constructed on a monthly time scale (predicted R2 values: 0.74-0.80); (3) The combination of phenology, 2-band enhanced vegetation index (EVI2), solarinduced chlorophyll fluorescence (SIF), killing degree days (KDD), and minimum vapor pressure deficit (VPDmin) with LSTM achieved the optimal yield estimation (RMSE = 740.10 kg/ha). Our findings demonstrated a scalable and effective method for predicting and estimating corn yield over a large area, which can potentially be applied to other crops in different geographical contexts.

Climate change and human activities are increasing the frequency, scale, and speed of forest insect damage (FID), threatening global forest health. However, satellite-based warning methods that rely solely on machine learning and driving factors remain limited by the complex mechanisms of FID and the poor quality of driver data. This study developed a hybrid warning approach that integrates spatiotemporal prior knowledge with factor-driven machine learning. For FID monitoring, we proposed a multi-index co-segmentation method to improve the accuracy of spatiotemporal information extraction. For FID warning, we extracted spatiotemporal prior knowledge from historical data—geographic constraints and an area prior derived from historical area dynamics—and incorporated them into a factor-driven model, which was evaluated across multiple years (2020–2023). The results demonstrate that this framework effectively mitigates the FID misclassification and structurally unreasonable severity allocation caused by purely factor-driven predictions. The main findings are as follows: (1) The co-segmentation-based FID identification achieved 89.3 % OA for disturbance detection and 86.6 % for disturbance classification. (2) In the primary 2023 experiment, the geographic constraint suppressed false alarms over healthy forest, while the area control further constrained the predicted infestation area to the range forecast from historical area dynamics, reducing commission error and raising the overall warning accuracy from 67.6 % to 69.1 % and the healthy-class recall from 0.83 to 0.97. Multi-year backcasting showed that the prior constraints consistently reduced total-area overprediction, but pixel-level accuracy improved only in 2023 and declined in earlier years when historical records were shorter. The framework should therefore be interpreted as a locally calibrated city-scale approach whose effectiveness depends on the quality and temporal stability of historical priors. These findings demonstrate that city-scale FID warning in the study area is constrained by the spatial scale and limited variability of driver data, and that spatiotemporal prior knowledge helps alleviate the limitations of factor-driven models.

Accurately estimating large-scale crop yields amid climate change is increasingly critical for crop management, marketing, and storage. Remote-sensing data-especially from moderate-resolution imaging spectroradiometer (MODIS)-are integral to such estimations, yet MODISs coarse spatial resolution causes notable mixed-pixel issues. Crop-specific masks can improve accuracy by filtering out nonagricultural pixels to obtain pure crop pixels, but the optimal mask percentage remains unclear. We evaluated six crop-mask thresholds (50%-100%) to define pure corn pixels for multiple MODIS-derived vegetation indices (VIs) [two-band enhanced vegetation index (EVI2)/normalized difference vegetation index (NDVI) from MOD09Q1 at 250 m/8-day; and enhanced vegetation index/NDVI from MOD09A1, MOD13Q1, and MOD13A1 at 500 m/8-day, 250 m/16-day, and 500 m/16-day], and integrated these VIs with regression and machine/deep-learning algorithms for county-level corn yield estimation across the U.S. Corn Belt. Results showed that the seasonal trajectories and interannual variabilities of VI-yield correlations derived from different mask thresholds were similar in both pattern and intensity. Importantly, mask threshold selection had a clear influence on corn yield estimation performance and was primarily governed by spatial resolution: for 250 m VIs, higher thresholds (80%-100%) generally achieved lower errors, whereas for 500 m VIs, lower thresholds (50%-60%) were more robust. In the extreme drought year (2012), 250 m VIs benefitted from higher thresholds and 500 m VIs achieved optimal performance at intermediate thresholds (70%-80%). These findings provide practical guidance for selecting crop-mask thresholds to improve MODIS-based large-scale corn yield estimation.

Timely and accurate estimation of crop area is fundamental for agricultural policy formulation, production forecasting, and economic assessment. Probability-sampling-based area estimation provides statistically rigorous inference with quantifiable uncertainty, but its operational implementation remains challenging. Two barriers are particularly important for probability-sampling-based in-season crop area estimation. First, collecting spatially dispersed probability samples in field is highly costly, especially in regions under complex traffic transportation conditions. Second, although progressively updated in-season crop maps provide valuable auxiliary information for improving estimation precision, standard post-stratified estimation is often impractical when samples were originally selected using an earlier stratification map (the updated post-strata may strongly overlap with the original sampling strata, leaving some substrata with insufficient samples). To address these two challenges, we present a practical framework for progressive in-season crop area estimation by introducing two complementary techniques. First, a vehicle–unmanned aerial vehicle (UAV) cooperative response design assigns sample units to multiple park-and-launch sites and minimizes the overall time cost of the vehicle tour and UAV sorties using ant colony optimization, thereby improving the efficiency of the probability samples collection. Second, a poststratified combined ratio estimator, a design-consistent estimation method that does not require sufficient samples within each substratum, is applied to progressively refine crop area estimates by leveraging the updated in-season crop maps in later season. In a field experiment conducted for rapeseed area estimation in Yanting County, 263 probability sample units selected under the initial stratification (the first in-season crop map) were collected in 31 hours using the proposed UAV cooperative sampling method. The poststratified combined ratio estimator was then introduced to update the crop area estimate using later-season crop maps, which reduced the standard error of the area estimate from 1.34% to 1.06% (≈21% relative reduction). Overall, the proposed framework provides a statistically robust rigorous and operationally feasible solution for progressive in-season crop area estimation.

Large-scale agricultural remote sensing monitoring is challenged by pronounced spatial heterogeneity arising from fragmented terrain, complex climatic backgrounds, and diverse cropping structures. However, existing agricultural zoning schemes generally lack an integrated consideration of remote sensing imaging mechanisms and key variable conditions such as atmospheric interference and crop phenology, limiting their direct utility in guiding region-specific sensor selection and classification algorithm calibration. To address this limitation, this study integrates multi-source earth observation data and agricultural statistical information to construct an Agricultural Remote-sensing Classification Difficulty Index (ARCDI) from multiple dimensions, including image availability, cropping structure, cropland fragmentation, and topographic environment. On this basis, a graph theory-based spatially constrained Skater clustering algorithm is introduced to establish a two-tier “cropland–major cereal crops” zoning framework oriented toward remote sensing applications. The results indicate that the proposed framework delineates five distinct first-tier cropland classification difficulty zones across China. This zoning scheme effectively quantifies the regional heterogeneities in monitoring challenges. Building upon this first-tier zoning, the framework is further refined into 50 second-tier major cereal crop classification difficulty zones, including 13 winter wheat zones, 21 maize zones, and 16 rice zones. Statistical tests and spatial analyses demonstrate that the proposed zoning scheme significantly outperforms conventional clustering approaches in balancing within-zone homogeneity and spatial continuity. This advantage is quantitatively reflected by consistently lower residual spatial autocorrelation (residual Moran’s I ≈ 0.10–0.11) and an approximately 20% reduction in within-zone variance compared with other spatially constrained methods. Extensive field-sample validation provides preliminary evidence of an inverse relationship between crop-type classification difficulty and accuracy. These results confirm the framework’s reliability in identifying regional difficulty and its decision-support value for selecting remote sensing strategies. Overall, this study systematically elucidates the spatial differentiation patterns of remote sensing classification difficulty for cropland and major cereal crops across China. The proposed framework provides robust scientific support for data selection, algorithm optimization, and differentiated strategy formulation in national-scale agricultural monitoring, thereby facilitating the operationalization of regional agricultural remote sensing applications.

Potato ranks as the third largest food crop in terms of global consumption volume and is crucial for food security and environmental sustainability. Early mapping aids in yield prediction and other preharvest decision-making. However, most early mapping methods rely heavily on current or historical crop samples, and their large-scale application is limited due to untimely collection or a lack of historical data. To address this challenge, we propose the potato early mapping index (PEMI), a method that does not require training samples; only Sentinel-2 (S2) multitemporal images were used to for early mapping of potato planting areas. The PEMI is designed to distinguish potatoes from other crops by multiplying the combination of red, red-edge 3 and NIR bands (related to spectral) during the greenness peak period with the spectral increment of red-edge 2 (related to phenology) from the sowing period to the greenness peak period. The effectiveness of the PEMI was validated in six typical potato planting regions (with different climates, cropping systems, and agricultural landscapes) in three major potato-producing countries worldwide (i.e., China, the USA, and France). The results indicate that, in the experimental area and five other validation areas from different years, the PEMI method using the natural breaks algorithm (automatic threshold method) generated potato maps two months before harvest (greenness peak period), with OAs ranging from 88.67% to 97.00% and F1 ranging from 79.57% to 94.12%, indicating that this method achieved early large-scale automatic mapping of potatoes. Compared with the post-season index (PL) using the empirical threshold, the PEMI OA and F1 values at Sites A (Idaho, the USA), C (Inner Mongolia Autonomous Region, China), E (Guangdong Province, China), and F (Hauts-de-France, France) exceeded those at PL, whereas at Sites B (Wisconsin, the USA) and D (Heilongjiang Province, China), they were lower. Overall, the PEMI outperforms the PL method. The PEMI method provides a robust and feasible solution for the early automatic identification of large-scale potatoes, carving a new path for crop early mapping and contributing to global food security forecasting.

Deep learning models have been widely used for vegetation destruction detection, but their performance depends on input features in addition to sample size and model architecture. Using Sentinel-2 imagery, we constructed multiple feature combinations from visible, near-infrared, and short-wave infrared bands and several spectral indices, and evaluated them with nine deep learning models. The results show that adding extra bands to RGB input improved most models, with RGBS yielding the most robust performance improvement among the tested band combinations, consistently outperforming the RGB baseline across the majority of deep learning architectures. Spectral indices further improved most models, especially BVDI, NBR, and SVI. Some single-index or two-index inputs achieved accuracy comparable to or higher than that of RGB input, indicating their potential for information compression and band substitution. However, model responses varied, and more input channels did not guarantee stable gains. Accuracy–efficiency analysis showed that RGBS, RGB + BVDI, RGB + NBR, and selected RGBNS + index combinations achieved favorable performance at low computational cost. Overall, optimized spectral inputs improve vegetation destruction detection, but their selection should balance model architecture, accuracy, and computational cost.

Satellites strive to strike a delicate balance between temporal and spatial resolution, thereby rendering the achievement of high resolution in both aspects challenging. Spatiotemporal fusion algorithms have emerged as a promising solution to tackle this challenge. However, with changes in spatiotemporal conditions, existing spatiotemporal fusion methods, particularly those based on deep learning, face challenges such as decreased prediction accuracy and poor reconstruction accuracy in areas of abrupt changes. This presents significant challenges for the fusion of multi-source remote sensing data to generate cloud-free remote sensing images on a daily scale. In this context, the study proposes a multiscale Attention-Guided deep optimization network for Spatiotemporal Data Fusion (AGSDF) method. The algorithm is designed to generate daily fine images using coarse image, based on historical reference fine images. Specifically, it firstly attempts to use a physical attention mechanism to mitigate the effects of climate change in time-series images. Implementing a continuous spatiotemporal fusion process across multiple scales significantly enhances the model's robustness. The performance of AGSDF was evaluated and compared to nine methods at six sites worldwide. The experimental results indicate that AGSDF achieved a top score in the assessment. Consequently, AGSDF holds high potential to produce accurate remote sensing products with high temporal and spatial resolution across extensive regions.

Precise, multi-scale mapping of agricultural landscapes-ranging from management-oriented Agricultural Parcels (APs) to fine-grained Crop Parcels (CPs)-is critical for land administration and precision agriculture. However, automated delineation remains challenging due to the accuracy limitations and poor generalizability of existing methods. While the Segment Anything Model (SAM) offers powerful zero-shot segmentation capabilities, its direct application is hindered by a lack of semantic awareness, a domain gap with RS imagery, and the prohibitive costs of full fine-tuning. To address these limitations, this study introduces PT-PEFT SAM, a novel framework that efficiently adapts SAM for agricultural mapping. A sequential "guide-then-specialize" adaptation strategy is employed, utilizing Prefix-Tuning (PT) to align SAM's representations with the remote sensing domain, followed by a proposed Orthogonalized Multi-Low-Rank Adaptation (Ortho-Multi-LoRA) to inject taskspecific semantics efficiently. The framework is further augmented by a learnable spectral adaptation module for multi-band compatibility and a domain-invariant feature learning module to enhance intrinsic generalization. Furthermore, a hierarchical architecture is introduced: the fine-tuned model performs semantic AP delineation, while a fully automated pipeline leveraging the pre-trained SAM 2 is deployed for intra-parcel CP segmentation, exploiting its superior zero-shot capabilities. Evaluations across European and Chinese datasets demonstrate that our method achieves state-of-the-art accuracy (Boundary F1 > 0.92 for APs; Geometric Error [GTC] <0.065 for CPs) and exceptional transfer learning performance. This research establishes a principled, scalable paradigm for adapting foundation models to specialized geospatial tasks, enabling the rapid and accurate generation of critical agricultural data.

Efficient detection of anomalous artificial surfaces is vital after disasters to support rapid emergency response. This study proposes a novel, training-free texture index method to accurately identify damaged artificial surfaces using post-disaster imagery. The method was tested across diverse disaster sites, including earthquake-affected areas in Turkey (Sites A-C), and tsunami- and tornado-damaged regions in Palu and Joplin (Sites D-I), covering a range of surface types and building structures. Sites A-C were chosen to develop and assess the efficacy of the proposed texture index method. Primarily, the three-dimensional texture features (3DTF) comprised of Contrast, Gabor wavelets, and secondary texture extraction (Con_Gabor), were amalgamated with a K-means classifier to delineate post-disaster artificial surface areas without prior knowledge. Given the discernible texture discrepancies between normal and anomalous artificial surfaces post-disaster, Homogeneity and Entropy texture features derived from Worldview-3 images at each site were leveraged to construct the artificial surface anomaly index (ASAI) for automatically extracting anomalous artificial surfaces. The findings demonstrated high overall accuracies of detecting anomaly artificial surfaces, ranging from 90.07 % to 91.76 % in Sites A-C. Notably, the ASAI outperformed Artificial Neural Network (ANN), Random Forest (RF), and the U-Net model. The overall accuracy (OA) of the ASAI method is 10 %-20 % higher than that of the U-Net model, showing superior automatic performance without necessitating training samples. Furthermore, the effectiveness of ASAI was validated in the Palu and Joplin sites, affirming its utility across diverse disaster scenarios. The damages of steel-tiled houses, middle-level houses and roads were effectively identified as anomaly artificial surfaces. The overall accuracies of the ASAI for identifying anomaly artificial surfaces were 89.55 %-92.5 %. These findings indicate the identification anomaly artificial surface using the ASAI was robust in different types of anomaly artificial surface caused by different disaster. The method of using ASAI to automatically identify anomaly artificial surfaces in post-disaster and single-temporal images has the potential for wide applicability.

The lack of ground-truth samples greatly limits the acquisition of spatiotemporally explicit crop distribution information essential for precision management, food security, and sustainable agricultural development. To address this limitation, various phenological knowledge-driven automated sample generation frameworks have been proposed. However, most studies prioritize the quantity and label accuracy of training samples while overlooking their spectral representativeness, which is crucial for balancing model accuracy and generalization performance. To tackle this challenge, this study proposed a simple index-based sample generation strategy that transformed crop mapping indices (i.e., phenology indices, PIs) into confidence indicators that characterize sample label accuracy, spectral representativeness, and spatial location, thereby guiding crop sampling and mapping. This approach satisfies the dual requirements of sample quantity and quality for model training while providing a practical solution to the limitations of index-based crop mapping in transfer applications. In this study, three well-validated PIs—Rice Phenology Index (RPI), Variation of the Vegetation-Pigment index (VVP), and Greenness and Water Content Composite Index (GWCCI)—were introduced to guide the sampling and mapping of rice, maize, and soybean in Northeast China from 2017 to 2025, with 2019 serving as the primary experimental year. First, a piecewise linear normalization procedure was developed based on threshold-based classification rules to scale the original PI values to a uniform scale of [-1, 1]. Subsequently, the relationships between the normalized PIs and the label accuracy, spectral properties, and within-field spatial locations of the corresponding samples were comprehensively evaluated. Then, crop sample subsets from different confidence levels were combined and fed into a Random Forest classifier for crop mapping to identify the optimal combination that balances model accuracy and generalization performance. Finally, the cross-temporal transferability of the proposed sampling strategy was further evaluated over extended time series. The results demonstrated that normalized PIs can serve as confidence indicators for guiding sample acquisition, allowing for an explicit assessment of label correctness and the distribution of spectral information in feature space. Classification results suggested that combining medium- and high-confidence samples ([0.25, 1]) yielded more satisfactory model accuracy (OA>90%) and performance (R2>0.91). Conversely, the results confirmed that relying solely on a single classification threshold is not advisable, as even a small proportion of non-random noises can substantially impair model training; similarly, relying only on super-high-confidence samples is also inadvisable. Consequently, by mitigating the detrimental effects of samples near the classification threshold, the proposed strategy effectively avoids direct classification errors when applied across spatiotemporal domains in index-based mapping methods, thereby enabling large-scale, long-term crop sampling and mapping. Ultimately, the study generated a large set of approximately 596,000 samples along with corresponding spatial distribution maps for rice, maize, and soybean across Northeast China from 2017 to 2025, achieving OA of 90.07%-98.60% and R2 of 0.8990-0.9877 (2019-2023). This study provides a general paradigm that utilizes emerging PIs to facilitate transfe

High-resolution Earth observation (EO) is critical for tracking spatially heterogeneous sustainable development goals (SDGs), such as cropland dynamics and urban expansion. However, persistent limitations in spatiotemporal continuity (satellite revisit gaps) and cost-efficiency (prohibitive pricing of commercial <2 m data) hinder its scalability. While deep learning-based single image super-resolution (SR) techniques offer a potential solution, their quantitative equivalence to native high-resolution data and the generalizability across geographies remain unproven. Here, we demonstrate that AI-powered SR can systematically transform freely available 10-m Sentinel-2 imagery with visible (RGB) and near infrared (NIR) bands into 2m-resolution images with RGB-NIR bands while preserving spectral-temporal fidelity. Specifically, we trained a geospatially constrained transformer-based SR framework (GeoSR) with 3.15 million km & sup2; of co-registered Gaofen-1/6 and Sentinel-2 pairs, achieving near-native performance with <1% F1-score loss in critical applications: cultivated land parcels mapping (F1-score = 0.84 vs. 0.85 for native Gaofen-1/6), urban footprint extraction (0.82 vs. 0.81), and fine-grained land use and land cover classification (0.43 vs. 0.43). Notably, the geospatial module enables large-scale generalization, retrospectively reconstructing decade-long environmental dynamics from historical archives - an unprecedented capability unattainable solely through launching new satellites. We further demonstrated operational scalability through two large-scale implementations: mapping 6.09 million hectares of agricultural parcels of the entire Anhui Province, China, and extracting 9866 km & sup2; of building footprints across 87 C40 coastal cities. As a free-to-access platform (www.sr-earth.org), GeoSR significantly reduces high-resolution data costs compared to commercial alternatives. AI-powered SR is not merely an image enhancement tool but a scientifically valid EO data source, particularly transformative for specific SDG monitoring in low/middle-income regions where native HR data scarcity impedes evidence-based policymaking.

Timely and accurate access to the spatial distribution of crops is critical for precision agriculture. However, large-scale automated in-season crop mapping often faces the challenge of limited training data. In this study, we proposed a dynamic and adaptive in-season winter wheat mapping (DAWM) framework, which leverages temporal-spectral similarity to minimize reliance on ground truth labels. The DAWM framework starts by utilizing historical prior knowledge from target domains to construct a temporal- spectral library for winter wheat, which integrates the phenological curves of winter wheat throughout the entire growing season and the spectral curves of winter wheat at different times. In this part, two modules—confidence learning (CL) and reclustering are used to refine historical crop maps and generate a reference temporal-spectral curve library. Then, a novel dynamic temporal-spectral similarity (DTSS) metric, proposed in this study, is used to produce training data for the random forest (RF) model by dynamically calculating the temporal-spectral similarity of each unlabeled objects to each cluster in the reference library. The proposed DAWM framework was evaluated in three global sites (10 000 km2 each) and compared with two existing phenology-based methods: time-weighted dynamic time warping (twDTW) and phenology-time weighted dynamic time warping (ptDTW). Results demonstrated that DAWM outperforms both methods, achieving F1-scores above 0.90 across all study sites and improving the F1-score by approximately 10% during the heading stage. In several study areas, DAWM also reached stable winter wheat identification about 45 days earlier than the compared methods, demonstrating its strong potential for dynamic in-season crop mapping. Extensive testing in the Huang-Huai-Hai Plain region of China achieved an F1 accuracy of 86.37%, while additional experiments in Heilongjiang Province demonstrated the method’s applicability for mapping maize and rice, with F1 accuracies of 90.35% and 96.31%, respectively. The proposed approach offers a robust solution for dynamic, high-quality crop mapping without relying on local ground labels, providing valuable data support for agricultural production management.

This study developed a 30-m resolution annual cropland dataset spanning 1988-2024 to resolve the unstable data quality and high sample acquisition costs in mapping cropland distributions in two agricultural regions of the Qinghai-Tibet Plateau (QTP): the Hehuang Valley (HV) and middle basin of the Yarlung Zangbo River and its two tributaries (the Lhasa and Nianchu rivers; MBYZR and LNR, respectively). This dataset was generated using Landsat imagery and training samples derived from visual interpretation. An initial classification was conducted using a Random Forest classifier. To ensure the stability of training sample quality across time, a sample cleaning approach was applied annually, based on spectral consistency constraints, allowing for the temporal extension of samples. The dataset demonstrated high classification accuracy, whereas the MBYZR and LNR demonstrated better classification performance, reflecting strong stability and robustness. Both regions showed favorable results regarding precision and recall, validating this approach's effectiveness in multi-temporal remote sensing classification. Therefore, this dataset provides critical support for cropland monitoring, food security assessment, and agricultural adaptation in QTP studies, offering a practical reference for time-series sample construction and transfer in remote sensing classification.

Remote sensing instance segmentation is a significant but difficult task due to the need for a number of accurate mask labels. Although existing weakly supervised methods use multiple points or bounding boxes for one object to reduce labeling requirements, they still require many manual labeling costs. Therefore, we propose an SPSIS model with only one point for each object to reduce the annotation burden. Pseudobounding boxes are firstly generated using the refined masks of the segment anything model and candidate points are randomly sampled within the boxes. In addition, we propose a point classification method combining ensemble learning and label propagation algorithm to classify sampled points. Finally, we use a point loss function so that the mask-based instance segmentation model can effectively adapt to the point samples. We have conducted extensive experiments on agricultural greenhouse and WHU datasets, demonstrating the superiority of SPSIS. In addition, SPSIS significantly lessens the precision difference between weakly and fully supervised instance segmentation