Luận án tiến sĩ: Time Series High Resolution Land Use Land Cover Mapping in Mainland Vietnam

Time Series High Resolution Land Use Land Cover Mapping in Mainland Vietnam January 2024 Truong Thinh Van Time Series High Resolution Land Use Land Cover Mapping in Mainland Vietnam A Dissertation Submitted to the Graduate School of Science and Technology, University of Tsukuba in Partial Fulfillment of Requirements for the Degree of Doctor of Philosophy in Environmental Studies Doctoral Program in Environmental Studies, Degree Programs in Life and Earth Sciences Truong Thinh Van Abstract Land use land cover (LULC) maps provide valuable information for understanding different anthropogenic-related processes, including climate change, urban expansion, urban heat island, and sediment disasters. In addition, LULC information is crucial due to the great impact of LULC changes to the Earth’s terrestrial ecosystem and the atmosphere. Given the importances of LULC data, the production of a time series of high-resolution and high-accuracy LULC maps in a large scale remains challenging due to lack of high temporal and high spatial resolution satellite data, lack of train- ing data, and time consuming. This study aims to investigate the LULC change in Vietnam via producing a time series of high-resolution LULC maps for mainland Viet- nam.

To achieve this purpose, I conducted several independent studies on addressing particular issues related to LULC mapping, including relationship between overall ac- curacy of LULC map and number of category (Chapter 2), studying the potential of multi-temporal synthetic aperture radar for mapping annual forest cover (Chapter 3), creation of high-resolution LULC map using a Time-Feature Convolutional Neural Net- work (CNN) (Chapter 4), and creation of a time series of high-resolution LULC maps (Chapter 5). These studies are independent but important puzzles to a comprehensive understanding about LULC mapping and LULC change assessment. A systematic review of relationship between overall accuracy and number of cat- egories was described in Chapter 2. By reviewing 64 papers on LULC mapping, a negative correlation was found between overall accuracy and number of categories.

This study suggested that increasing the number of LULC categories by one could result in about 0.77% decrease in overall accuracy. The result suggested that the pro- cess of selecting number of LULC category should also consider the expected overall accuracy of output LULC map. In addition, the number of category and accuracy of a LULC map should be both considered in the assessment on the map’s reliability. A study on mapping annual forest cover for Vietnam was described in Chapter 3.

By employing multi-temporal PALSAR-2/ScanSAR data, the result showed that annual forest cover maps of Vietnam achieved much higher overall accuracy (86.1%) than that of JAXA (77. PALSAR-2/ScanSAR, by containing seasonal infor- mation, is more reliable than single-temporal mosaic PALSAR-2 data. Therefore, it is more robust for annual forest mapping than the single-temporal data. The result of analyzing annual change of forest area revealed an increasing trend of forest cover in Vietnam from 2015-2018, which is consistent with the statistical data of the Vietnamese government.

Given the potential of PASAR-2/ScanSAR data, this study decided to use it in combination with other satellite images to produce a high-resolution LULC map in Chapter 4. The product of a high-resolution LULC map for Vietnam was presented in Chapter 4. By modifying a novel CNN approach proposed by Hirayama et al. (2022), this study used satellite data with 6 seasons instead of 4 seasons as in the original approach.

Using a rigorous validation dataset by stratified random sampling method, the result LULC map achieved high overall accuracy of 90.2% for 12 categories , which is higher than existing 10-m LULC map for Vietnam. In addition, this study proposed a simple and practical approach for making a time series of cloud-free satellite image using multi-temporal data from different years. This approach is effective for removing cloud and bad pixel and can easily to be implemented in Google Earth Engine. Based on the results in Chapter 4, the production of a time series of LULC maps from 2017-2022 was described in Chapter 5.

A data migration method was applied to transfer the reference data from 2020 to other years from 2017 to 2022. The preliminary assessment using migrated reference data showed an overall accuracy greater than 94% for all maps from 2017 to 2022, surpassing the benchmark value of 89.2% for the LULC map with 14 category. Furthermore, existing LULC maps of Vietnam showed that they have accuracy lower/equal their benchmark value suggested using the result in Chap- ter 2. This study’s LULC maps outperformed existing 10-m LULC maps of Vietnam in term of spatial resolution and overall accuracy.

This study revealed that the rate of forest loss in Vietnam might be 4 times higher than national statistical data, although forest area in Vietnam have experienced a net increase during 2017-2022. This finding was supported by other study, which showed that the forest loss caused by rubber plantation in Southeast Asia should be at least 2-3 time higher than expected. This study is the first attempt to provide a time series of high-accuracy 10-m LULC maps for Vietnam with 14 category, including a new category of solar panel. The result of this study is expected to support the diverse applications such as LULC change analy- sis, disaster related studies, biodiversity conservation, planning, food security, carbon emission and climate modelling, solar energy planning.

ii Acknowledgement First, I gratefully acknowledge the financial support of the Project for Human Resource Development Scholarship by Japanese Grant Aid (JDS) for providing scholarship and valuable supports to my entire studying time in Japan. I also acknowledge VNU De- velopment at Hoa Lac campus for allowing me to do my Ph. I would like to thank my chief supervisor, Dr. NASAHARA Nishida Kenlo, for his guidance and support.

I have learned valuable fundamental knowledge related to remote sensing and land use land cover mapping from him. Moreover, I admire his philosophy of teaching students and doing research. From him, I have learnt that doing research is not a duty but a hobby and curiosity is necessary to a researcher. I very much appreciate my co-supervisors, Dr.

MATSUSHITA Bunkei, Dr. TSU- JIMURA Maki and the rest of the dissertation committees, Dr. UCHIDA Taro, Dr. YAMAKAWA Yosuke, Dr.

ASANUMA Jun, Dr. YABAR Mostacero Helmut Friedrich, and Dr. MIZUNOYA Takeshi who have given thoughtful questions and constructive feedback to improve this study. I appreciate support from the Japan Aerospace Exploration Agency and the Remote Sensing Technology Center of Japan, especially Mr.

HIRAYAMA Sota for sharing the source code of SACLASS-2, Dr. TADONO Takeo, Dr. HAYASHI Masato, and Ms. OH- GUSHI Fumi for data provision and publishing the High Resolution Land Use Land Cover Map Products of Vietnam.

I express my thanks to Dr. PHAN Cao Duong and Dr. HOANG Thanh Tung for helping to collect reference data across Vietnam. I sincerely appreciate support from my current organization (Center for VNU Development at Hoa Lac campus) in Vietnam, especially Mr.

NGUYEN Huu Hieu (current director), Mrs. NGUYEN Thi Hue (current vice director), Dr. TRAN Duc Dang (former director) for their tremendous support and positive encouragement. I acknowledge all the laboratory members, who gave me valuable comments and support during my Ph.

I very much appreciate my families for their love and supports, especially my par- ents for doing hard to secure my life. Finally, I would like to express my special thanks to my wife and my daughter, who have supported me and become the biggest motivation during my studying period in Japan. 11 Contents Abstract i Acknowledgement iii List of Figures Vii List of Tables xi List of Publications xii List of Acronyms and Abbreviations xiv 1 Introduction 1 11 Background of the research problem.2 The status of global LULC mapping .3 The need of having a time series of high resolution LULC maps for Vietnam 4 1.4 Method of LULC classification. 5 15 Accuracy assessment of LULC map.6 Challenge and opportunities for LULC mapping.

6 1/7 Research aims and objectives. ee es 7 2 The relationship between accuracy of land use and land cover map with number of map’s category 2. ee ee ee 8 2. ee ee eee 9 2.1 Land cover map’s accuracy in individual studies.2 Land cover map’s accuracy synthesized from multiple studies.

Q Q Q Q Q ee 14 Annual Forest Cover Maps for Vietnam During 2015-2018 Using ALOS-2/PALSAR- 2 and Auxiliary Data 15 3.2 Satellite Images and Preprocessing.5 Method for Accuracy Assessment and Area Estimation .6 Method for Comparing FNF Maps from Difference Sources .1 Classification result for threetestsies.2 Forest/Non-Forest Maps Using ScanSAR/NDVI/SRIM. Comparison of Forest Cover Data from Multiple Sources .1 Improvement of Forest Classification Using Multi-Temporal ScanSAR Images. Q Q Q Q HQ HH HH ko 36 3.2 Forest Change in Vietnam Between 2015 and 2018 .3 Forest Dynamicsin Vietnam .4 The Need of Having Reliable Forest Maps for User Communities 39 3. eee 39 Creation of a High-Resolution Land Use Land Cover Map for Vietnam Using a Time-Feature Convolutional Neural Network 41 4.2 Satellite images pre-processing.

Reference data collection and accuracy assessment.1 Time series ofSentinel2images.2 LULC map of mainland Vietnam.3 Comparison with other LULC products. Q Q Q ee n g ng ng k k Và 65 4. The practicability of utilizing multi-year data for cloud-free opti- cal image generation. ee ee ee 67 4.3 Multi-temporal data for LULC classification .4 Potentials of time-feature CNN for future studies.

ee 68 5 The first time series of 10-m land use and land cover maps at the national scale for Vietnam from 2017 to 2022 70 5.2 Creation of a time series of reference data and accuracy assessment 71 5.1 A time series of LULC maps for Mainland Vietnam .2 LULC change in Vietnam from 2017to2022.3 Updating LULC map forVietnam.1 Comparing accuracy of different LULC product using the bench- mark study .2 Support to national statisticaldata.3 Application for future LULC predicion.4 The demand on high-resolution LULC maps of Vietnam .5 Limitation and futurestudies. ee 91 6 Conclusions and Remarks 93 6.1 Summary of key findings and contributions.2 Limitations and recommendations. 94 Bibliography 95 Mái List of Figures 2.1 The variation of overall accuracy with different number of class in single studies (a) The study of Herold et al. (2008); (b) The study of McCombs etal.2 The correlation between overall accuracy and the number of LULC classes (for 64 studies including MODIS and Landsat).3 The average accuracy with respect to sensor types (the error bars are the standard deviation).

Q eee ee ee eee 14 3.1 Examples of the input data used to create forest/non-forest (FNF) maps. (a) False-color composite of ScanSAR images for mainland Vietnam (dates: Jun. 20, 2017): red: HH polarization, green: HV polarization, blue: difference of HH and HV polarizations. (b) MODIS NDVI image for mainland Vietnam (date: Nov.

(c) Slope (in degree) images of mainland Vietnam estimated using SRTM- DEM images (date: Feb.2 The work-flow used to produce FNF maps for mainland Vietnam using multi-temporal PALSAR-2/ScanSAR, multi-temporal MODIS NDVI, and SRTM images.3 Distribution of reference data. (a) Distribution of training data. (b) Dis- tribution of the 500 validation sample points for the mainland Vietnam based on the 2017 FNF map. (c, d, e) Distribution of the 400 validation sample points based on the 2017 ScanSAR/NDVI/SRTM FNF map at three test sites.4 Classification result of FNF maps for three test sites in Vietnam using: (a) Location of the test sites in mainland Vietnam, (b) single-temporal PALSAR-2 mosaic images, (c) multi-temporal PALSAR-2/ScanSAR im- ages, and (d) a combination of multi-temporal PALSAR-2/ScanSAR, MODIS NDVI, and SRTM-slope images.5 Classification result for FNF maps for mainland Vietnam from 2015 to 2018.

(a) Forest/non-forest map in 2015, (b) Forest/non-forest map in 2016, (c) Forest/non-forest map in 2017, (d) Forest/non-forest map in 2018.6 Comparison of the overall accuracy of FNE maps from a different data source (Error bars are standard errors).7 Comparison of the forest fraction of different forest cover maps for 2015 with a spatial resolution of 1 km.