Temporal Fusion Research Articles

Learning U-Net without forgetting for near real-time wildfire monitoring by the fusion of SAR and optical time series

Wildfires are increasing in intensity and frequency across the globe due to climate change and rising global temperature. Development of novel approach to Monitor wildfire progressions in near real-time is therefore of critical importance for emergency responses. The objective of this research is to investigate continuous learning with U-Net by exploiting both Sentinel-1 SAR and Sentinel-2 MSI time series for increasing the frequency and accuracy of wildfire progression mapping. In this study, optical-based burned areas prior to each SAR acquisition (when available) were accumulated into SAR-based pseudo progression masks to train a deep residual U-Net model. Unlike multi-temporal fusion of SAR and optical data, the temporal fusion of progression masks allows us to track as many wildfire progressions as possible. Specifically, two approaches were investigated to train the deep residual U-Net model for continuous learning: 1) Continuous joint training (CJT) with all historical data (including both SAR and optical data); 2) Learning without forgetting (LwF) based on newly incoming data alone (SAR or optical). For LwF, a mean squared loss was integrated to keep the capabilities learned before and prevent it from overfitting to newly incoming data only. By fusing optical-based burned areas, SAR-based progression pseudo masks improve significantly, which benefits both data sampling and model training, considering the challenges in SAR-based change extraction attributed to the variability in SAR backscatter of the surrounding environments. Pre-trained ResNet was frozen as the encoder of the U-Net model, and the decoder part was trained to further refine the derived burned area maps in a progression-wise manner. The experimental results demonstrated that LwF has the potential to match CJT in terms of the agreement between SAR-based results and optical-based ground truth, achieving a F1 score of 0.8423 on the Sydney Fire (2019–2020) and 0.7807 on the Chuckegg Creek Fire (2019). We also observed that the SAR cross-polarization ratio (VH/VV) shows good capability in suppressing multiplicative noise and detecting burned areas when VH and VV have diverse temporal behaviors.

A Deep Multi-Frame Super-Resolution Network for Dynamic Scenes

The multi-frame super-resolution techniques have been prosperous over the past two decades. However, little attention has been paid to the combination of deep learning and multi-frame super-resolution. One reason is that most deep learning-based super-resolution methods cannot handle variant numbers of input frames. Another reason is that it is hard to capture accurate temporal and spatial information because of the misalignment of input images. To solve these problems, we propose an optical-flow-based multi-frame super-resolution framework, which is capable of dealing with various numbers of input frames. This framework enables to make full use of the input frames, allowing it to obtain better performance. In addition, we use a spatial subpixel alignment module for more accurate subpixel-wise spatial alignment and introduce a dual weighting module to generate weights for temporal fusion. Both two modules lead to more effective and accurate temporal fusion. We compare our method with other state-of-the-art methods and conduct ablation studies on our method. The results of qualitative and quantitative analyses show that our method achieves state-of-the-art performances, demonstrating the advantage of the designed framework and the necessity of proposed modules.

Temporal Fusion Approach for Video Classification with Convolutional and LSTM Neural Networks Applied to Violence Detection

In modern smart cities, there is a quest for the highest level of integration and automation service. In the surveillance sector, one of the main challenges is to automate the analysis of videos in real-time to identify critical situations. This paper presents intelligent models based on Convolutional Neural Networks (in which the MobileNet, InceptionV3 and VGG16 networks had used), LSTM networks and feedforward networks for the task of classifying videos under the classes "Violence" and "Non-Violence", using for this the RLVS database. Different data representations held used according to the Temporal Fusion techniques. The best outcome achieved was Accuracy and F1-Score of 0.91, a higher result compared to those found in similar researches for works conducted on the same database.

Hierarchical Temporal Attention Network for Thyroid Nodule Recognition Using Dynamic CEUS Imaging.

Contrast-enhanced ultrasound (CEUS) has emerged as a popular imaging modality in thyroid nodule diagnosis due to its ability to visualize vascular distribution in real time. Recently, a number of learning-based methods are dedicated to mine pathological-related enhancement dynamics and make prediction at one step, ignoring a native diagnostic dependency. In clinics, the differentiation of benign or malignant nodules always precedes the recognition of pathological types. In this paper, we propose a novel hierarchical temporal attention network (HiTAN) for thyroid nodule diagnosis using dynamic CEUS imaging, which unifies dynamic enhancement feature learning and hierarchical nodules classification into a deep framework. Specifically, this method decomposes the diagnosis of nodules into an ordered two-stage classification task, where diagnostic dependency is modeled by Gated Recurrent Units (GRUs). Besides, we design a local-to-global temporal aggregation (LGTA) operator to perform a comprehensive temporal fusion along the hierarchical prediction path. Particularly, local temporal information is defined as typical enhancement patterns identified with the guidance of perfusion representation learned from the differentiation level. Then, we leverage an attention mechanism to embed global enhancement dynamics into each identified salient pattern. In this study, we evaluate the proposed HiTAN method on the collected CEUS dataset of thyroid nodules. Extensive experimental results validate the efficacy of dynamic patterns learning, fusion and hierarchical diagnosis mechanism.

Attention-Based Parallel Multiscale Convolutional Neural Network for Visual Evoked Potentials EEG Classification.

Electroencephalography (EEG) decoding is an important part of Visual Evoked Potentials-based Brain-Computer Interfaces (BCIs), which directly determines the performance of BCIs. However, long-time attention to repetitive visual stimuli could cause physical and psychological fatigue, resulting in weaker reliable response and stronger noise interference, which exacerbates the difficulty of Visual Evoked Potentials EEG decoding. In this state, subjects' attention could not be concentrated enough and the frequency response of their brains becomes less reliable. To solve these problems, we propose an attention-based parallel multiscale convolutional neural network (AMS-CNN). Specifically, the AMS-CNN first extract robust temporal representations via two parallel convolutional layers with small and large temporal filters respectively. Then, we employ two sequential convolution blocks for spatial fusion and temporal fusion to extract advanced feature representations. Further, we use attention mechanism to weight the features at different moments according to the output-related interest. Finally, we employ a full connected layer with softmax activation function for classification. Two fatigue datasets collected from our lab are implemented to validate the superior classification performance of the proposed method compared to the state-of-the-art methods. Analysis reveals the competitiveness of multiscale convolution and attention mechanism. These results suggest that the proposed framework is a promising solution to improving the decoding performance of Visual Evoked Potential BCIs.

Video-based driver action recognition via hybrid spatial–temporal deep learning framework

Driver action recognition aims to distinguish normal driver action and some abnormal driver actions such as leaving the wheel, talking on the phone, diving with smoking, etc. For the purpose of traffic safety, studies on the computer vision technologies for driver action recognition have become especially meaningful. However, this issue is far from being solved, mainly due to the subtle variations between different driver action classes. In this paper, we present a new video-based driver action recognition approach based on the hybrid spatial–temporal deep learning framework. Specifically, we first design an encoder–decoder spatial–temporal convolutional neural network (EDSTCNN) to capture short-term spatial–temporal representation of driver actions jointly with optical flow prediction. Second, we exploit the feature refinement network (FRN) to refine the short-term driver action feature. Then, convolutional long short-term memory network (ConvLSTM) is employed for long-term spatial–temporal fusion. Finally, the fully connected neural network (FCNN) is used for final driver action recognition. In our experiment, we validate the performance of the proposed framework on our self-created datasets, including a simulated driving dataset and a real driving dataset. Extensive experimental results illustrate that the proposed hybrid spatial–temporal deep learning framework obtains the highest accuracy in multiple driver action recognition datasets (98.9% on SEU-DAR-V1 dataset and 97.0% on SEU-DAR-V2 dataset).

Temporal Fusion Transformer with 285 Explanatory Variables Outperforms ARIMAX in Predicting Next-Day Emergency Department Overcrowding: A Retrospective Study

Temporal Fusion Transformer with 285 Explanatory Variables Outperforms ARIMAX in Predicting Next-Day Emergency Department Overcrowding: A Retrospective Study

A Simple Spatio–Temporal Data Fusion Method Based on Linear Regression Coefficient Compensation

High spatio–temporal resolution remote sensing images are of great significance in the dynamic monitoring of the Earth’s surface. However, due to cloud contamination and the hardware limitations of sensors, it is difficult to obtain image sequences with both high spatial and temporal resolution. Combining coarse resolution images, such as the moderate resolution imaging spectroradiometer (MODIS), with fine spatial resolution images, such as Landsat or Sentinel-2, has become a popular means to solve this problem. In this paper, we propose a simple and efficient enhanced linear regression spatio–temporal fusion method (ELRFM), which uses fine spatial resolution images acquired at two reference dates to establish a linear regression model for each pixel and each band between the image reflectance and the acquisition date. The obtained regression coefficients are used to help allocate the residual error between the real coarse resolution image and the simulated coarse resolution image upscaled by the high spatial resolution result of the linear prediction. The developed method consists of four steps: (1) linear regression (LR), (2) residual calculation, (3) distribution of the residual and (4) singular value correction. The proposed method was tested in different areas and using different sensors. The results show that, compared to the spatial and temporal adaptive reflectance fusion model (STARFM) and the flexible spatio–temporal data fusion (FSDAF) method, the ELRFM performs better in capturing small feature changes at the fine image scale and has high prediction accuracy. For example, in the red band, the proposed method has the lowest root mean square error (RMSE) (ELRFM: 0.0123 vs. STARFM: 0.0217 vs. FSDAF: 0.0224 vs. LR: 0.0221). Furthermore, the lightweight algorithm design and calculations based on the Google Earth Engine make the proposed method computationally less expensive than the STARFM and FSDAF.

Self-adaptive combination method for temporal evidence based on negotiation strategy

In temporal information fusion, the information collected by sensors is obtained dynamically with the passage of time. Unlike the spatial information fusion, temporal fusion should be dynamic. Evidence theory has been applied to fuse temporal and spatial information; however, existing temporal fusion methods do not treat conflicting and non-conflicting evidence sources distinctively. Moreover, unlike spatial evidence sources, which are obtained simultaneously, temporal evidence sources cannot be evaluated simultaneously to obtain their degree of reliability. Thus, it is necessary to develop a method for temporal evidence combination. In this paper, a self-adaptive combination method for temporal evidence is proposed based on the negotiation strategy. In the proposed method, a set called an evidence set is constructed by the cumulative temporal fusion results of the previous moment, current moment, and future moment. The evidence set is evaluated as conflicting or non-conflicting according to the maximum power pignistic probability distance between each pair of evidence sources in the set. Then, temporal evidence sources are self-adaptively combined by different methods according to the degree of conflict. Numerical experiments were conducted to evaluate the performance of the proposed method. The results indicated that the proposed method is sufficiently effective and robust to support decision making.

Recursive Multi-Signal Temporal Fusions With Attention Mechanism Improves EMG Feature Extraction

The design of pattern recognition-based myoelectric interfaces has been heavily explored and contested in the research literature. A considerable proportion of the performance of these interfaces has been linked to the quality of the feature extraction (FE) stage used to describe the underlying myoelectric signal. In this paper, we address two important factors of FE that have not been fully exploited; 1) Spatial focus - traditional FE methods focus mainly on the concatenation of features extracted from individual channels and 2) Temporal focus - available FE methods are cross-sectional in nature, largely ignoring the temporal information that may exist between feature windows. To overcome these limitations, several spatiotemporal FE methods have been proposed including hand-crafted and deep learning (DL) models, with the latter showing significant performance enhancements at the cost of increased computational burden. This paper tackles the aforementioned limitations by 1) proposing novel extensions to simple time-domain FE methods, including the waveform length, zero crossings and root mean square, that can capture the relation between any number of channels, and 2) leveraging the long short-term memory concepts from deep neural networks to build a recursive framework that overcomes the cross-sectional nature of traditional methods. The advantages offered by the proposed Recursive Multi-Signal Temporal Fusion (RMTF) features include improved performance, competing with state-of-the-art FE methods without the computational costs associated with leading DL models, and the simplicity of the concepts making them suitable for real-time implementations. Experiments on 65 intact-limbed and amputee subjects reveal an approximate average of 15% reduction in classification errors as compared to models built with other feature sets.

Monitoring cyanobacterial harmful algal blooms at high spatiotemporal resolution by fusing Landsat and MODIS imagery

Toxic Cyanobacteria-rich Harmful Algal Blooms (CyanoHABs) are severe environmental issues impacting water environment, aquatic life populations, and surrounding wild and human lives. Given their fast-changing variations in space and time, frequent monitoring of CyanoHABs with sufficient spatial details and coverage is needed to understand their impacts. However, current monitoring activities based on in situ or satellite data do not meet this requirement due to the limited spatial coverages of buoy measurements and the compromise between spatial resolution and temporal frequency of satellite observations. In this study, we develop a Spatial-Temporal Image Fusion (STIF) approach to enable high spatial-temporal resolution monitoring of CyanoHABs. The proposed approach consists of two steps: (1) a new CyanoHAB spectral index called Broad Wavelength Algae Index (BWAI) is developed for fine-but-sparse (Landsat) and coarse-but-frequent (MODIS) satellite images, and (2) the Landsat and MODIS derived BWAI images are fused by the Robust Adaptive Spatial and Temporal Fusion Model (RASTFM) to generate fine-and-frequent Landsat-like BWAI images. Our results show that the proposed BWAI index is with higher similarity and correlation with the reference Cyanobacteria Index (CI) images and in situ observations than the comparative algae indices devised for broad wavelength sensors. Moreover, the 30-m Landsat-like BWAI image series provide more accurate and detailed results than their 500-m MODIS counterparts and greatly improve the temporal frequency over Landsat-based algae indices, demonstrating the contribution of the improved spatial-temporal resolution achieved by the proposed STIF approach. Consequently, this research fills the gap of high spatiotemporal resolution monitoring of CyanoHABs and paves a new way of assessing water environment in a timely and detailed manner.

Harmonious attention network for person re-identification via complementarity between groups and individuals

Person re-identification (Re-ID) is of important capability for artificial intelligence and human–computer interaction. The main challenge of person Re-ID lies in limited data to precisely capture a wide range of appearance variations over multiple viewpoints. Furthermore, compared to the Re-ID between the single person, the person groups contain information about the relationship between pedestrians that can potentially help identify certain identities. The Re-ID combining groups and individuals remain to be a promising task under rare study. In this paper, we propose a harmonious attention network for person re-identification, in which we jointly consider the complementarity between person groups and individuals. Concretely, first we propose a two-stream attentive network (TSAN) to respectively learn the information from the person groups and individuals. TSAN consists of a spatial–temporal fusion network for the group Re-ID, as well as a deep network for the traditionally individual person Re-ID. To jointly consider the contributions of the groups and individuals, then we propose a novel re-ranking algorithm (GIRK) based on the learned features to associate the group and individual information. We also propose a new group Re-ID dataset DukeGroupVid to evaluate the performance of our approach. Comprehensive experimental results on the proposed dataset and other Re-ID datasets demonstrate the effectiveness of our model.

Video super-resolution based on a spatio-temporal matching network

Deep spatio-temporal neural networks have shown promising performance for video super-resolution (VSR) in recent years. However, most of them heavily rely on accuracy motion estimations. In this paper, we propose a novel spatio-temporal matching network (STMN) for video super-resolution, which works on the wavelet domain to reduce dependence on motion estimations. Specifically, our STMN consists of three major components: a temporal fusion wavelet network (TFWN), a non-local matching network (NLMN), and a global wavelet domain residual connection (GWDRC). TFWN adaptively extracts temporal fusion wavelet maps via three 3d convolutional layers and a discrete wavelet transform (DWT) decomposition layer. The extracted temporal fusion wavelet maps are rich in spatial information and knowledge of different frequencies from consecutive frames, which are feed to NLMN for learning deep wavelet representations. NLMN integrates super-resolution and denoising into a unified module by pyramidally stacking non-local matching residual blocks (NLMRB). At last, GWDRC reconstructs the super-resolved frames from the deep wavelet representations by using global wavelet domain residual information. Consequently, our STMN can efficiently enhance reconstruction quality by capturing different frequencies wavelet representations in consecutive frames, and does not require any motion compensation. Extensive experiments conducted on benchmark datasets demonstrate the effectiveness of our method compared with state-of-the-art methods.

Mapping wetland using the object-based stacked generalization method based on multi-temporal optical and SAR data

Wetland ecosystems have experienced dramatic challenges in the past few decades due to natural and human factors. Wetland maps are essential for the conservation and management of terrestrial ecosystems. This study is to obtain an accurate wetland map using an object-based stacked generalization (Stacking) method on the basis of multi-temporal Sentinel-1 and Sentinel-2 data. Firstly, the Robust Adaptive Spatial Temporal Fusion Model (RASTFM) is used to get time series Sentinel-2 NDVI, from which the vegetation phenology variables are derived by the threshold method. Subsequently, both vertical transmit-vertical receive (VV) and vertical transmit-horizontal receive (VH) polarization backscatters (σ0 VV, σ0 VH) are obtained using the time series Sentinel-1 images. Speckle noise inherent in SAR data, resulting in over-segmentation or under-segmentation, can affect image segmentation and degrade the accuracies of wetland classification. Therefore, we segment Sentinel-2 multispectral images to delineate meaningful objects in this study. Then, in order to reduce data redundancy and computation time, we analyze the optimal feature combination using the Sentinel-2 multispectral images, Sentinel-2 NDVI time series, phenological variables and other vegetation index derived from Sentinel-2 multispectral images, as well as time series Sentinel-1 backscatters at the object level. Finally, the stacked generalization algorithm is utilized to extract the wetland information based on the optimal feature combination in the Dongting Lake wetland. The overall accuracy and Kappa coefficient of the object-based stacked generalization method are 92.46% and 0.92, which are 3.88% and 0.04 higher than that using the pixel-based method. Moreover, the object-based stacked generalization algorithm is superior to single classifiers in classifying vegetation of high heterogeneity areas.

Rethinking Temporal Fusion for Video-Based Person Re-Identification on Semantic and Time Aspect

Recently, the research interest of person re-identification (ReID) has gradually turned to video-based methods, which acquire a person representation by aggregating frame features of an entire video. However, existing video-based ReID methods do not consider the semantic difference brought by the outputs of different network stages, which potentially compromises the information richness of the person features. Furthermore, traditional methods ignore important relationship among frames, which causes information redundancy in fusion along the time axis. To address these issues, we propose a novel general temporal fusion framework to aggregate frame features on both semantic aspect and time aspect. As for the semantic aspect, a multi-stage fusion network is explored to fuse richer frame features at multiple semantic levels, which can effectively reduce the information loss caused by the traditional single-stage fusion. While, for the time axis, the existing intra-frame attention method is improved by adding a novel inter-frame attention module, which effectively reduces the information redundancy in temporal fusion by taking the relationship among frames into consideration. The experimental results show that our approach can effectively improve the video-based re-identification accuracy, achieving the state-of-the-art performance.

Robustly detect different types of text in videos

Text in videos can be categorized into three types: overlaid text, layered text, and scene text. Existing detection methods focus on a specific type of text and cannot obtain a good performance when working on other text types. To our knowledge, few works explore to build a system to simultaneously detect all types of text. In this paper, we present a unified video text detector, which can simultaneously localize all types of text in videos accurately. Our system consists of a spatial text detector and a temporal fusion filter. First, we explore to use three different strategies to learn the spatial text detector based on deep convolutional neural networks, so that it can simultaneously detect various texts without knowing the types of text. Then, a new area-first non-maximum suppression computation combined with multiple constraints is proposed to remove the redundant bounding boxes. Finally, the temporal fusion filter exploits the features of spatial locations and text components to integrate the detection results of consecutive frames to further remove false positives. To validate the proposed approach, comprehensive experiments are carried out on three publicly available datasets, consisting of overlaid text, layered text, and scene text. The experimental results demonstrate that our method consistently achieves the best performance compared with state-of-the-art methods.

Monitoring of Sugarcane Growth Based on the Fused Remote Sensing NDVI Series and Ground Seeding Survey

HighlightsThe spatial and temporal fusion model ESTARFM was used to obtain NDVI timing data with high fusion accuracy and high spatial and temporal resolution.High-quality NDVI timing data could be obtained by using ESTARFM to fuse HJ-1 CCD and MODIS data.Fused NDVI data coupled with ground seeding survey data could effectively monitor sugarcane growth status.Abstract. This study addressed the instability of clear-sky remote sensing data with high spatial resolution in sugarcane growing areas in southern China and the current inconsistency between traditional survey results and remote sensing results for seedling growth. Moderate-resolution imaging spectroradiometer (MODIS) data and China land resources satellite (HJ-1 CCD) data were used to build high-resolution normalized difference vegetation index (NDVI) time series using the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM). Agronomic indicators of sugarcane were obtained by field sampling and were used for determining the remote sensing monitoring index (NDVI) of sugarcane growth. The method provided satisfactory results for evaluating sugarcane growth, with accuracy exceeding 90%. Moreover, sugarcane growth monitoring in a wider area was highly correlated with yield per unit area. Keywords: Growth status, HJ-1 CCD, MODIS, NDVI time series, Spatial and temporal fusion, Sugarcane.

ST-DeepHAR: Deep Learning Model for Human Activity Recognition in IoHT Applications

Human activity recognition (HAR) has been regarded as an indispensable part of many smart home systems and smart healthcare applications. Specifically, HAR is of great importance in the Internet of Healthcare Things (IoHT), owing to the rapid proliferation of Internet of Things (IoT) technologies embedded in various smart appliances and wearable devices (such as smartphones and smartwatches) that have a pervasive impact on an individual’s life. The inertial sensors of smartphones generate massive amounts of multidimensional time-series data, which can be exploited effectively for HAR purposes. Unlike traditional approaches, deep learning techniques are the most suitable choice for such multivariate streams. In this study, we introduce a supervised dual-channel model that comprises long short-term memory (LSTM), followed by an attention mechanism for the temporal fusion of inertial sensor data concurrent with a convolutional residual network for the spatial fusion of sensor data. We also introduce an adaptive channel-squeezing operation to fine-tune convolutional a neural network feature extraction capability by exploiting multichannel dependency. Finally, two widely available and public HAR data sets are used in experiments to evaluate the performance of our model. The results demonstrate that our proposed approach can overcome state-of-the-art methods.

Identification of Sugarcane with NDVI Time Series Based on HJ-1 CCD and MODIS Fusion

It is currently difficult to acquire the clear-sky data with high spatial resolutions in spring and summer in the southern region of China, making it impossible to carry out timely and fine monitoring of sugarcane planting information. Thus, Fusui, a sugarcane producing county in Guangxi, was selected in this paper to analyze the NDVI characteristics and change rules during the whole growth period of sugarcane based on MODIS and HJ-1 CCD remote sensing data, which were fused into 30 m resolution NDVI time series data with high accuracy by using the spatial and temporal fusion model of multi-source remote sensing data ESTRAFM. In addition, the NDVI change rate and sample automatic training threshold were used to construct the sugarcane planting information identification model. The results showed that the fused images showed a high similarity with the observed images, indicating good fusion quality. Moreover, the correlation coefficients in the sugarcane planting area reached 0.953, and AD, AAD and SD were 0.033, 0.019 and 0.007, respectively. The NDVI change rate model was used to identify the sugarcane planting information in different time phases of 113 d, 129 d, 145 d, 193 d and 209 d in spring and summer, and the overall accuracy was 92.17%, 92.58%, 91.78%, 90.52% and 91.17%, respectively. The established model also achieved good results in 2017 with the overall accuracies are 88.44%, 87.79%, 89.79%, 88.34% for 113 d, 145 d, 193 d and 209 d.

A Two-Stage Fusion Framework to Generate a Spatio–Temporally Continuous MODIS NDSI Product over the Tibetan Plateau

The Tibetan Plateau (TP) is an important component of the global environmental system, on which the snow cover greatly affects the regional climate and ecology. Moderate resolution imaging spectroradiometer (MODIS) snow cover products have been demonstrated to be appropriate for investigating the snow cover over the TP. However, they are subject to cloud obscuration, and the TP’s extremely complex terrain makes the snow monitoring difficult. Therefore, in this paper, we propose a two-stage spatio–temporal fusion framework for the cloud removal of MODIS C6 snow products, including an adjusted Terra and Aqua combination (TAC) and a spatio–temporal fusion based on Gaussian kernel function and error correction (STF-GKF-EC). To the best of our knowledge, this is the first time that a spatio–temporally continuous daily 500-m MODIS normalized difference snow index (NDSI) product has been generated for the TP, which greatly improves the spatial and temporal resolutions of the current snow cover products. The main stage, STF-GKF-EC, adaptively weights the spatial and temporal correlations by the Gaussian kernel function, and further takes the rapid changes of snow cover into consideration through the error correction. The experiments indicated that STF-GKF-EC removes clouds completely, achieving an overall accuracy (OA) and mean absolute error (MAE) of 91.48% and 3.88, respectively. Based on the cloud-removed results, during 2001–2017, as far as the intra-annual variation is concerned, a large proportion of the snow cover appears between October and May, with a peak in February/March, and the variation is mainly controlled by temperature. For the inter-annual variation, an obvious increasing trend of 0.68/year for NDSI is observed before 2005, followed by a slight decreasing trend of 0.16/year, in which precipitation is a better explanation factor than temperature.