Since agricultural drought plays a leading role in restricting agricultural productivity, accurate forecasting is crucial for agricultural management. Although soil moisture (SM) is the primary variable for identifying and forecasting agricultural drought, accurately predicting SM is challenged by its strong interaction with external meteorological forcings that change rapidly across space and time. To provide a predictive method that reduces uncertainty in SM modeling and compensates for the latency in satellite-based SM products, we propose a meteorologically driven Transformer framework (MDTF). The framework can predict global surface (0–5 cm) and root-zone (0–100 cm) SM utilizing meteorological forecasts and soil physical properties with a prediction latency of 35 days. When validated against satellite-based SM data and in situ measurements, our proposed model more accurately predicts global SM spatial patterns and seasonal dynamics compared with the physics-based Common Land Model (CoLM), demonstrating consistent performance across different land covers. MDTF outperforms popular machine learning models, achieving unbiased root-mean-square error (ubRMSE) values of 0.0297 m3/m3 and 0.0211 m3/m3 for surface and root-zone SM. The predicted seamless global daily SM information could be effectively utilized for agricultural drought forecasting. Our research demonstrates that the MDTF model has unique advantages in modeling key hydrological variables of the Earth system, providing a reference for predicting time series of global SM and agricultural drought dynamics.

The leaf area index (LAI) is a basic vegetation structural variable in the ecosystem and climate model. Predicting vegetation LAI time series is essential for understanding the response of vegetation growth processes to climate change. Current predictive numerical phenology models are still subject to large uncertainties. Developing LAI time series prediction models is beneficial to improve the ability of land surface models to predict future vegetation growth. This study develops a deep learning model based on meteorological variables and attentional mechanisms to predict vegetation LAI time series. We tested our ALSTM for different biome types across North America and evaluated the model using satellite and flux tower data. The results demonstrated that the ALSTM outperforms other baseline models. The predicted ALSTM LAI shows high temporal and spatial consistency compared to satellite-based LAI. The ALSTM model can effectively predict downstream gross primary productivity (GPP) when coupled with the common land model (CoLM). ALSTM performs high accuracy in predicting the LAI of terrestrial vegetation and provides a reference for developing biophysical methods to predict future vegetation ecological processes for land surface models.

Vegetation phenology is a key indicator of ecosystem processes and biosphere feedbacks to the climate system. Remote sensing techniques have been widely used for large scale phenology observations. In this study, the first-order derivative (FOD) method was used to extract the start of the growing season (SOS) and the end of the growing season (EOS) based on the GIMMS LAI3g time series from 1982 to 2015. The results showed that the global 34-year mean SOS and EOS exhibited significant latitudinal variation in spatial distribution. In the Northern Hemisphere, SOS and EOS showed insignificant trend (1982–2015). In the Southern Hemisphere, EOS showed significantly delayed trend of 0.1251 days/year. In addition, GIMMS SOS and EOS showed a significant positive correlation on a global scale with MODIS SOS and EOS. Overall, we produced global vegetation phenology data with longer time series (1982–2015) compared to MODIS vegetation dynamics products (MCD12Q2).

Building change detection from very high-spatial-resolution (VHR) remote sensing images has gained increasing popularity in a variety of applications, such as urban planning and damage assessment. Detecting fine-grained “from–to” changes (change transition from one land cover type to another) of buildings from the VHR images is still challenging as multitemporal representation is complicated. Recently, fully convolutional neural networks (FCNs) have been proven to be capable of feature extraction and semantic segmentation of VHR images, but its ability in change detection is untested and unknown. In this letter, we leverage the semantic segmentation of buildings as an auxiliary source of information for the fine-grained “from–to” change detection. A deep multitask learning framework for change detection (MTL-CD) is proposed for detecting building changes from the VHR images. MTL-CD adopts the encoder–decoder architecture and solves the main task of change detection and the auxiliary tasks of semantic segmentation simultaneously. Accordingly, the change detection loss function is constrained by the auxiliary semantic segmentation tasks and enables the back-propagation of the building footprints’ detection errors for the improvement of change detection. A building change detection data set named the Guangzhou data set is also developed for model evaluation, in which the bitemporal R–G–B images were collected by airplane (2009) and unmanned aerial vehicle (UAV, 2019) with different flight heights. Experiments on the Guangzhou data set demonstrate that the MTL-CD method effectively detects fine-grained “from–to” changes and outperforms the postclassification methods and the direct change detection methods.

Tradeoffs among the spatial, spectral, and temporal resolutions of satellite sensors make it difficult to acquire remote sensing images at both high spatial and high temporal resolutions from an individual sensor. Studies have developed methods to fuse spatiotemporal data from different satellite sensors, and these methods often assume linear changes in surface reflectance across time and adopt empirical rules and handcrafted features. Here, we propose a dense spatiotemporal fusion (DenseSTF) network based on the convolutional neural network (CNN) to deal with these problems. DenseSTF uses a patch-to-pixel modeling strategy that can provide abundant texture details for each pixel in the target fine image to handle heterogeneous landscapes and models both forward and backward temporal dependencies to account for land cover changes. Moreover, DenseSTF adopts a mapping function with few assumptions and empirical rules, which allows for establishing reliable relationships between the coarse and fine images. We tested DenseSTF in three contrast scenes with different degrees of heterogeneity and temporal changes, and made comparisons with three rule-based fusion approaches and three CNNs. Experimental results indicate that DenseSTF can provide accurate fusion results and outperform the other tested methods, especially when the land cover changes abruptly. The structure of the deep learning networks largely impacts the success of data fusion. Our study developed a novel approach based on CNN using a patch-to-pixel mapping strategy and highlighted the effectiveness of the deep learning networks in the spatiotemporal fusion of the remote sensing data.

Detecting changes using bitemporal remote sensing imagery is vital to understand the dynamics of the land surface. Existing change detection models based on deep learning suffer from the problem of scale variation and pseudochange due to their insufficient multilevel aggregation and inadequate capability of feature representation, which limits the accuracy. This study proposes a densely attentive refinement network (DARNet) to improve change detection on bitemporal very-high-resolution remote sensing images. DARNet is based on the U-shape encoder–decoder architecture with the Siamese network as a feature extractor. The dense skip connection module (DSCM) is employed between the decoder and the encoder to aggregate multilevel feature maps. The hybrid attention module (HAM) is integrated to exploit contextual information and generate discriminative features. The recurrent refinement module (RRM) is exploited to progressively refine the predicted change maps during the decoding process. Experiments on testing the model performance were conducted on three benchmark datasets: the season-varying change detection (SVCD) dataset, the Sun Yat-sen University change detection (SYSU-CD) dataset, and the Learning Vision and Remote Sensing Laboratory building change detection (LEVIR-CD) dataset. The experimental results demonstrate that DARNet outperforms state-of-the-art models with kappa of 96.58%, 75.35%, and 90.69% for the SVCD, SYSU-CD, and LEVIR-CD datasets, respectively.

Vegetation phenology is a significance for studying the impact of climate change on ecosystems. Most of the current methods use a single characteristic variable to indirectly measure the vegetation phenology. So the accuracy of the phenological measurement will change due to the accuracy of the measurement index and many other phenological information have not been discovered. In this study, we evaluated leaf phenology using the convolutional neural regression network (CNNR) at a single and multiple sites in deciduous broad-leaved forest (DBF) from PhenoCam cameras. The error of the recognition result of the same site is about 3 days, the R2 of all stations was 0.843, and the error of the RMSE result is about 25 days. These finding provide a certain contribution to the research of vegetation phenology on the scale of daily time detection results.

Automatic extraction of building is of great importance in many remote sensing applications. Nevertheless, existed deep-learning based methods tend to generate building footprints with fuzzy boundaries and the geometric details could not be well-preserved. In this paper, we propose a bottom-up deep learning based approach for automatic building polygon reconstruction. Specifically, an end-to-end learnable model, multitask DeepLabv3+, is designed for predicting building footprint, corner heatmap and edge heatmap. Then, A geometric-based strategy is utilized to reconstruct the extracted corners as a closed polygon based on the edge heatmap. By doing so, our method is able to generate polygons much closer to real buildings with fine-grained boundaries and well-preserved geometric details. Compared to state-of-the-art models, experiments on high-resolution aerials imagery dataset demonstrate that our proposed approach achieves better performance in quantitative and qualitative results.

Normalized difference vegetation index (NDVI) carries valuable information related to the photosynthetic activity of vegetation and is essential for monitoring phenological changes and ecosystem dynamics. The medium to high spatial resolution satellite images from Landsat 8 and Sentinel 2 offer opportunities to generate dense NDVI time series at 10-m resolution to improve our understanding of the land surface processes. However, synergistic use of Landsat 8 and Sentinel 2 for generating frequent and consistent NDVI data remains challenging as they have different spatial resolutions and spectral response functions. In this letter, we developed an attentional super resolution convolutional neural network (ASRCNN) for producing 10-m NDVI time series through fusion of Landsat 8 and Sentinel 2 images. We evaluated its performance in two heterogeneous areas. Quantitative assessments indicated that the developed network outperforms five commonly used fusion methods [i.e., enhanced deep convolutional spatiotemporal fusion network (EDCSTFN), super resolution convolutional neural network (SRCNN), spatial and temporal adaptive reflectance fusion model (STARFM), enhanced STARFM (ESTARFM), and flexible spatiotemporal data fusion (FSDAF)]. The influence of the method selection on the fusion accuracy is much greater than that of the fusion strategy in blending Landsat-Sentinel NDVI. Our results illustrate the advantages and potentials of the deep learning approaches on satellite data fusion.

Monitoring crop areas is a key issue in remote sensing studies. A Crop Proportion Phenology Index (CPPI) model has previously been developed for estimation of winter wheat areas. Here we test the CPPI model in different areas using remote sensing data for varied kernel functions, including linear regression (LR), Artificial Neural Network (ANN), and Support Vector Regression (SVR). The differences of the model performances among different kernel functions were found to be small for areas with simple planting structure. For areas where multiple crop types have similar phenology cycles, the non-linear model of ANN was found to perform the best. This study indicates that the CPPI model can be applied to map winter wheat distribution in areas with complex planting structures, thus it holds promises for estimating fractional areas of winter wheat areas over large geographic areas.