Long Time Series Land Cover Classification in China from 1982 to 2015 Based on Bi-LSTM Deep Learning

Wang, Haoyu; Zhao, Xiang; Xin, Zhang; Wu, Deming; Du, Xiangdong

doi:10.3390/rs11141639

Cited by 72 publications

(29 citation statements)

References 80 publications

(89 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…SVM is widely considered as a powerful technique for classification tasks, while RF has some advantages such as short training time, easy parameterization, and high robustness to high-dimensional input features [13,64]. Concerning deep learning models, we selected three approaches that performed well in previous SITS classification studies: CNN-1D (1-dimensional CNN) [21], LSTM [22,23], and Bidirectional LSTM (Bi-LSTM) [65,66]. These architectures have been proven to have advantages to capture temporal dependencies in sequential data.…”

Section: A Evaluation Criteria and Methodsmentioning

confidence: 99%

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Yuan¹,

Liu²

2020

Preprint

View full text Add to dashboard Cite

<div>Satellite image time series (SITS) classification is a major research topic in remote sensing and is relevant for a wide range of applications. Deep learning approaches have been commonly employed for SITS classification and have provided state-of-the-art performance. However, deep learning methods suffer from overfitting when labeled data is scarce. To address this problem, we propose a novel self-supervised pre-training scheme to initialize a Transformer-based network by utilizing large-scale unlabeled data. In detail, the model is asked to predict randomly contaminated observations given an entire time series of a pixel. The main idea of our proposal is to leverage the inherent temporal structure of satellite time series to learn general-purpose spectral-temporal representations related to land cover semantics. Once pre-training is completed, the pre-trained network can be further adapted to various SITS classification tasks by fine-tuning all the model parameters on small-scale task-related labeled data. In this way, the general knowledge and representations about SITS can be transferred to a label-scarce task, thereby improving the generalization performance of the model as well as reducing the risk of overfitting. Comprehensive experiments have been carried out on three benchmark datasets over large study areas. Experimental results demonstrate the effectiveness of the proposed method, leading to a classification accuracy increment up to 2.38% to 5.27%. The code and the pre-trained model will be available at https://github.com/linlei1214/SITS-BERT upon publication.</div><div><b>This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.</b></div>

show abstract

Section: A Evaluation Criteria and Methodsmentioning

confidence: 99%

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Yuan¹,

Liu²

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…For traditional machine learning algorithms, we selected SVM and RF classifiers. Concerning deep learning network models, we selected two approaches that performed well in previous SITS classification studies: long short-term memory (LSTM) network [19,20], and Bidirectional LSTM (Bi-LSTM) network [61,62]. These LSTM-based architectures have advantages over other deep learning models to capture long-term temporal dependences in sequence data [63], and thus they are suitable to be used as benchmarks.…”

Section: A Evaluation Criteria and Methodsmentioning

confidence: 99%

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Yuan¹,

Liu²

2020

Preprint

View full text Add to dashboard Cite

Satellite image time series (SITS) classification is a major research topic in remote sensing and is relevant for a wide range of applications. Deep learning approaches have been commonly employed for SITS classification and have provided state-of-the-art performance. However, deep learning methods suffer from overfitting when labeled data is scarce. To address this problem, we propose a novel self-supervised pre-training scheme to initialize a Transformer-based network by utilizing large-scale unlabeled data. In detail, the model is asked to predict randomly contaminated observations given an entire time series of a pixel. The main idea of our proposal is to leverage the inherent temporal structure of satellite time series to learn general-purpose spectral-temporal representations related to land cover semantics. Once pre-training is completed, the pre-trained network can be further adapted to various SITS classification tasks by fine-tuning all the model parameters on small-scale task-related labeled data. In this way, the general knowledge and representations about SITS can be transferred to a label-scarce task, thereby improving the generalization performance of the model as well as reducing the risk of overfitting. Comprehensive experiments have been carried out on three benchmark datasets over large study areas. Experimental results demonstrate the effectiveness of the proposed method, leading to a classification accuracy increment up to 1.91% to 6.69%. <div><b>This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.</b></div>

show abstract

“…Previous works indicated great potential in the utilization of non-linear deep learning methods for land cover classification with time series data from EO multispectral systems [13,14,22]. The Sentinel-2 time series together with the combined reference dataset for 2017 were inputs for the supervised classification algorithm in this study.…”

Section: Generation Of Annual Land Cover Dynamic Mapsmentioning

confidence: 99%

Improved Estimations of Nitrate and Sediment Concentrations Based on SWAT Simulations and Annual Updated Land Cover Products from a Deep Learning Classification Algorithm

Samarinas

Tziolas

Zalidis

2020

IJGI

View full text Add to dashboard Cite

The agricultural sector and natural resources are heavily interdependent, comprising a coherent but complex system. The soil and water assessment tool (SWAT) is widely used in assessing these interdependencies for regional watershed management. However, long-term simulations of agricultural watersheds are considered as not realistic since they have often been performed assuming constant land use over time and are based on the coarse resolution of the existing global or national data. This work presents the first insights of the synergy among SWAT model and deep learning classification algorithms to provide annually updated and realistic model’s parameterization and simulations. The proposed hybrid modelling approach couples the physical process SWAT model with the versatility of Earth observation data-driven non-linear deep learning algorithms for land use classification (Overall Accuracy (OA) = 79.58% and Kappa = 0.79), giving a strong advantage to decision makers for efficient management planning. A validation case at an agricultural watershed located in Northern Greece is provided to demonstrate their synergistic use to estimate nitrate and sediment concentrations that load in Zazari Lake. The SWAT model has been implemented under two different simulations; one with the use of a static coarse land use map and the other with the use of the annual updated land use maps for three consecutive years (2017–2019). The results indicate that the land use changes affect the final estimations resulting to an enhanced prediction performance of 1% and 2% for sediment and nitrate, respectively, when the annual land use maps are incorporated into SWAT simulations. In this context, a hybrid approach could further contribute to addressing challenges and support a data-centric scheme for informed decision making with regard to environmental and agricultural issues on the river basin scale.

show abstract

Long Time Series Land Cover Classification in China from 1982 to 2015 Based on Bi-LSTM Deep Learning

Cited by 72 publications

References 80 publications

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Self-Supervised Pre-Training of Transformers for Satellite Image Time Series Classification

Improved Estimations of Nitrate and Sediment Concentrations Based on SWAT Simulations and Annual Updated Land Cover Products from a Deep Learning Classification Algorithm

Contact Info

Product

Resources

About