Haijun Li

Haijun Li

PhD Candidate

Department of Geographical Science

University of Maryland

Posts by Collection

fieldwork

Poyang Lake fieldwork

Jiangxi, China, 2014-09

In-situ measurements of water quality parameters including chlorophyll concentration, suspended sediments concentration, water clarity, total nitrogen (TN) and total phosphorus (TP) in Poyang Lake.

Taizhou Bay fieldwork

Zhejiang, China, 2016-08

Collecting data including chlorophyll concentration, suspended sediments concentration, water clarity in Taizhou Bay, Zhejiang.

Poyang Lake fieldwork

Jiangxi, China, 2017-03

Collecting soil moisture measurements data from hydrological stations around Poyang Lake.

GLAD crop survey in the Southern Great Plains

US, 2022-06

Collecting ground information of crop type and crop phenology over stratified random sites for maize, soybean and winter wheat maps validation, in the states of NE, KS, OK, CO, NM and TX.

GLAD crop survey in the southeast US

US, 2023-08

Collecting ground information of crop type and crop phenology over stratified random sites for maize and soybean maps validation, in the states of MD, PA, WV, OH, IN, IL, MS, KY, TN, AR, MS, AL, LA, GA, SC, NC and VA.

GLAD crop survey in the upper Midwest

US, 2025-08

Collecting ground information of crop type and crop phenology over stratified random sites for corn, soybean and wheat maps validation, in the states of ND, SD, MT, MN.

portfolio

publications

Development of a 10-m resolution maize and soybean map over China: Matching satellite-based crop classification with sample-based area estimation

Published in Remote Sensing of Environment, 2023

Li, H., Song, X.P., Hansen, M.C., Becker-Reshef, I., Adusei, B., Pickering, J., Wang, L., Wang, L., Lin, Z., Zalles, V. and Potapov, P. (2023). Development of a 10-m resolution maize and soybean map over China: Matching satellite-based crop classification with sample-based area estimation. Remote sensing of environment, 294, p.113623.
Download Paper

talks

Development of a 10 m Resolution Maize and Soybean Map Over China

Published:

Spatially explicit crop information is essential for food security and agricultural sustainability. However, high-resolution crop type maps are unavailable in most countries. In this study, we developed an experimental workflow and produced the first 10 m resolution maize and soybean map over China. To support crop mapping, we implemented a stratified, two-stage, cluster sampling design and collected ground data over the entire country in 2019. We developed a multi-scale, multi-temporal procedure for mapping, in which field data were used as training to map maize and soybean over the first-stage samples (10 km × 10 km equal-area blocks) with PlanetScope and Sentinel-2 data. Then, the classified blocks were used as training to map maize and soybean over the country with wall-to-wall Sentinel-2 data and random forests. We used all available Sentinel-2 surface reflectance data acquired between April and October, 2019 and created monthly image composites as inputs for random forests. We derived in-season maize and soybean area estimates using the statistical field survey data and a regression estimator. Utilizing the probability output layer of random forests models, we found and applied empirical probability thresholds that matched map-based and sample-based area estimates. Maize area was estimated to be 330,609 ± 34,109 km2, and soybean area was estimated to be 78,107 ± 12,969 km2. Validated using the field samples as reference, our crop type map had an overall accuracy of 91.8 ± 1.2%. The user’s and producer’s accuracy for the maize class were 93.9 ± 2.5% and 79.2 ± 3.6%, respectively, and for the soybean class were 63.6 ± 12.1% and 61.9 ± 11.8%, respectively. Our map-based maize and soybean area estimates had close agreements with government reports at the provincial and prefectural levels, with r2 of 0.90 and 0.92 for maize, and 0.93 and 0.94 for soybean, respectively. Our experimental workflow can generate harmonized results for crop area estimation and crop type mapping simultaneously, within the growing season, and independent from government statistics. As Sentinel-2 data are being acquired consistently and very-high-resolution commercial satellite data are increasingly available, our established workflow may be applied in an operational setting for annual crop type mapping in China and other countries.

Li, H., Song, X.P., Hansen, M.C., Becker-Reshef, I., Adusei, B., Pickering, J., Wang, L., Wang, L., Lin, Z., Zalles, V., Potapov, P., Stehman, S.V. & Justice, C.O. (2022). Development of a 10 m Resolution Maize and Soybean Map Over China. AGU Fall Meeting, Abstract GC23A-03, December 11-15, Chicago, Illinois, USA.

Crop Type and Yield Mapping Using Long-term Satellite Observations, Weather and Field Data

Published:

Long-term spatially explicit information on crop production is essential for understanding food security in a changing climate. Land used to produce soybean is rapidly expanding in the agricultural frontiers of South America, replacing natural vegetation, pasture, and other cropland. In this research, we mapped annual soybean cover in South America between 2000 and 2020 by combining field data and satellite observations from Landsat, Sentinel-2 and MODIS. Validated against statistical field data, our soybean cover maps had an overall accuracy of >95%. We then developed a multi-temporal, multi-scale modeling workflow for mapping soybean yield in Brazil. Our method integrated the soybean cover maps, municipality-level crop yield statistics, Landsat and MODIS data, climate and weather records, and random forests regression. Compared to the reference data from official statistics, our annual soybean yield maps had an overall RMSE of 418 kg/ha and an r2 of 0.60. We show that the crop classification and yield models trained on historical data could be used to produce crop type and yield maps for future years, demonstrating the predictive capability of our approach for operational crop monitoring applications.

Song, X. P., Li, H., Potapov, P., & Hansen, M., (2022). Crop Type and Yield Mapping Using Long-term Satellite Observations, Weather and Field Data. AGU Fall Meeting, Abstract B43B-01, December 11-15, Chicago, Illinois, USA.

Testing global stilling with ERA5 reanalysis data

Published:

“Global Terrestrial Stilling”, a noticeable decrease in wind speeds around the globe, has been reported, with implications for energy generation and evaporation. The two leading possible causes are increased surface roughness due to land use changes and general climate changes. However, some long-term regional studies show no trend of decreasing wind speeds. Global wind speed data since 1979 is available in reanalysis datasets. Here, we use ERA5 to map decadal averages of 10m wind speed. The results show that, when comparing the 1980s and 2010s, there is a complex pattern of small increases (within 5% of zero), with notable additions (10 to 20%) scattered within equatorial regions and decreases (5 to 10%) in polar regions and portions of southern Asia, Europe, and elsewhere. However, from the 1980s to the 2010s, there was a significant decrease in global wind speed by 1.82%.

Rahman S.S., Lee J.A., Li H., Aziz M.T. (2023). Testing global stilling with ERA5 reanalysis data. AAG Annual Meeting, March 23-27, Denver, Colorado, USA.

Decadal Changes in Relative Aeolian Transport Potential in Major Global Dust Source Regions

Published:

Lee J.A., Aziz M.T., Li H., Wang X. (2023). Decadal Changes in Relative Aeolian Transport Potential in Major Global Dust Source Regions. 2023 International Conference on Aeolian Research, July 9-14, Las Cruces, New Mexico, USA.

Lee J.A., Aziz M.T., Li H., Wang X. (2023). Decadal Changes in Relative Aeolian Transport Potential in Major Global Dust Source Regions. 2023 International Conference on Aeolian Research, July 9-14, Las Cruces, New Mexico, USA.

10-m Crop Mapping Using Satellite Data, Field Survey and Machine Learning over North America

Published:

Spatially explicit crop maps are essential for agricultural management and food security. However, timely producing high-resolution crop maps over large areas with diverse crop types and heterogeneous fields remains a challenge. To date, only a few operational crop mapping programs exist, such as the Cropland Data Layer program in the US and the Annual Crop Inventory program in Canada. Although these 30-m maps have overall accuracies above 85% for major crops, they are produced and delivered publicly several months after harvest and have relatively lower accuracies over fragmented landscapes. In this study, we used an operational workflow to produce 10-m maps for maize and soybean over North America, from 2019 to 2022. From April to November each year, we downloaded Sentinel-2 surface reflectance data and corrected the bidirectional reflectance distribution function (BRDF) effects to generate nadir BRDF-adjusted reflectance. We then created 10-day composites and gridded the data in non-overlapping 1° × 1° tiles. We implemented temporal linear interpolation on the pixel basis to fill the data gaps in the composites. We collected a large amount of field data as reference and trained multi-year random forests models for maize and soybean, respectively. We validated the crop maps using a probability field sample as reference. Results showed that our map in 2022 achieved an overall accuracy of 95.5 ± 0.4%. The user’s and producer’s accuracy for maize were 93.7 ± 1.5% and 82.6 ± 2.4%, and for soybean were 91.8 ± 1.9% and 84.7 ± 1.8%, respectively. Our developed workflow can produce high-resolution crop maps at the continental scale. We anticipate that this framework is applicable for operational annual crop mapping in other large areas.

Li, H., Song, X.P., Adusei, B., Pickering, J. and Hansen, M. (2023). 10-m Crop Mapping Using Satellite Data, Field Survey and Machine Learning over North America. AGU Fall Meeting, Abstract GC53H-0902, December, San Franscico, California, USA.

Advancing 10-m Crop Mapping Using All Sentinel-2 Observations over the Contiguous United States

Published:

Li, H., Song, X.P., Adusei, B., Pickering, J., Lima, A., Poulson, A.J., Baggett, A., Potapov, P., Stehman, S.V. and Hansen, M. (2024), Advancing 10-m Crop Mapping Using All Sentinel-2 Observations over the Contiguous United States. AGU Fall Meeting, Abstracts GC13O-0407, December 9-13, Washington, D.C., USA.

Li, H., Song, X.P., Adusei, B., Pickering, J., Lima, A., Poulson, A.J., Baggett, A., Potapov, P., Stehman, S.V. and Hansen, M. (2024), Advancing 10-m Crop Mapping Using All Sentinel-2 Observations over the Contiguous United States. AGU Fall Meeting, Abstracts GC13O-0407, December 9-13, Washington, D.C., USA.

Mapping multiple crops of multiple seasons: a 30 m national crop map of Paraguay derived from the Harmonized Landsat and Sentinel-2 (HLS) data and stratified random sampling

Published:

Crop maps are essential data for a wide range of applications such as crop growth monitoring, crop production forecasting and natural disaster assessments. However, many countries lack the capacity to produce annual crop updates. The Harmonized Landsat and Sentinel 2 (HLS) product, with its 30-m spatial resolution and 2–3-day revisit frequency, offers an unprecedented opportunity for agricultural remote sensing. We implemented an end-to-end operational workflow for estimating crop area and mapping crop type over the country of Paraguay, where multiple cropping seasons with diverse crop varieties exist during a calendar year. We developed an automated satellite data processing system on High-Performance Clusters (HPC) to convert satellite images into grided Analysis Ready Data (ARD) products, which were subsequently used as input for machine learning-based crop mapping. Based on the field sample, we estimated the first-season soybean area in Paraguay over 2022/2023 to be 2.96 million hectares (Mha) with a standard error (S.E.) of 0.20 Mha, the second-season soybean area to be 0.57 ± 0.13 Mha, the maize area to be 1.26 ± 0.10 Mha and the rice area to be 0.17 ± 0.04 Mha S.E. We produced two 30 m resolution crop classification maps for the two cropping seasons that matched the sample-based area estimates. The overall accuracies of the two maps were both 98% with high users’ and producer’s accuracies. Our results demonstrate that HLS data in conjunction with stratified random sampling are adequate to map multiple crops of multiple seasons on the national scale. This cost-effective method shows substantial potential for operational implementation by government agencies.

Song, X.P., Zalles, V., Adusei, B., Pickering, J., Hernandez-Serna, A., Li, H., Torales, E., Miranda, L., Stehman, S.V., Caballero, P. and Hansen, M. (2024). Mapping multiple crops of multiple seasons: a 30 m national crop map of Paraguay derived from the Harmonized Landsat and Sentinel-2 (HLS) data and stratified random sampling. AGU Fall Meeting, Abstract B11G-1389, December 9-13, Washington, D.C., USA.

Defining and Mapping Tropical Dryland Woody Cover: A Comparative Study using Drone Data in Senegal

Published:

Dryland woody cover plays an important role in terrestrial ecosystems while undergoing transformation due to the changing climate and human impacts. Dryland woody cover mapping is thus critical but challenging given the great variations in dryland vegetation life forms, structure, and spatial distribution. Thus, the clear thresholds defining the structural limits of woody vegetation cover class are needed, to map woody cover and to validate the map accuracy and usability. The Land Cover Classification System (LCCS) of Food and Agriculture Organization (FAO) provides a detailed operational framework for applying classifiers in building legends for mapping land cover. The physiognomic-structural traits - height and cover, have been used by a number of global classification schemes to define the extent of woody cover. However, in the existing studies on dryland woody cover mappings, the definitions of “tree” or “forest” are often vague or lacking in terms of definitive physiognomic structural perspectives of woody cover. The objective of this study is to apply classifiers in mapping the woody cover, and compare with existing large-scale products to illustrate the need for clear definitions. The study was implemented in Senegal which features a corresponding high variation in height and cover of woody vegetation. We present spatially explicit woody cover structure mapping using field-based unmanned aerial vehicles (UAV) of passive optical photogrammetry and lidar-derived point cloud. We employed Harmonized Landsat Sentinel-2 (HLS) and Synthetic Aperture Radar (SAR) data from Sentinel-1 as mapping inputs. Specifically, we generated a set of 30-m cover maps for woody cover derived using different height definitions. We validated the cover maps using set aside field data and achieved a set of R-square ranging from 0.67 to 0.86, which demonstrates the capability of HLS and SAR in mapping different definitions of woody cover. We also validated the existing large-area maps to examine what structural thresholds these products best match using the aerial lidar collections. The study emphasizes the need to relate remote sensing-based observations and derived products with the explicit structure of woody cover, illustrating that imprecise definitions preclude robust validation and the informed use of such products.

Li, X., Hansen, M., Potapov, P., Li, H. and Poulson, A.J. (2024). Defining and Mapping Tropical Dryland Woody Cover: A Comparative Study using Drone Data in Senegal. AGU Fall Meeting, Abstract B11J-1445, December 9-13, Washington, D.C., USA.

Detecting Oil and Gas Land Dynamics in the Permian Basin by Using Deep Learning and Very High-resolution Aerial Images

Published:

The Permian Basin is one of the most significant oil and gas (O&G)-producing regions in the United States. The exploitation of O&G resources in this region has been going on for over a hundred years. Although the fossil energy output supports a large part of the country and contributes significantly to the national economy, there are several negative consequences associated with exploitation, including greenhouse gas emissions, biodiversity loss, ground instability and environmental pollution. Detecting O&G land and monitoring land change dynamics are crucial for mitigating these negative effects and managing resources effectively. In this research, we adopted a Convolutional Neural Network (CNN) based deep learning model to detect the O&G land in the Permian Basin from 2010 to 2018 using very high-resolution aerial images from the National Agriculture Imagery Program (NAIP). An O&G energy land change map is also produced from the classification results. Our prototype map of 2018 demonstrates that an accurate semantic segmentation algorithm over different land cover types at a 0.6 m spatial resolution is possible on super-computing clusters. Our results, anchored by a 2018 prototype, reveal significant changes in the extent and use of oil and gas (O&G) lands, demonstrating robustness in challenging environments such as bare ground, developed regions, and shrublands where the O&G land can be easily misclassified with the background objects. This sets the foundation for future time series analyses to further understand the dynamic nature of O&G activities across various regions. Our research offers the potential to generate an up-to-date map of O&G land change over large regions at a very high resolution, which can assist the government and stakeholders in making informed decisions and managing the resources accordingly.

Ma, Y., Song, X.P., Hernandez Serna, W.A., Li, H., Lu, Z. and Silva, J. (2024). Detecting Oil and Gas Land Dynamics in the Permian Basin by Using Deep Learning and Very High-resolution Aerial Images. AGU Fall Meeting, Abstract B33B-1546, December 9-13, Washington, D.C., USA.

teaching

GEOG371-0101: Programming for Image Analysis

Teaching Assistant, Undergraduate course, Department of Geographical Sciences, University of Maryland, 2023

The course provides an introduction to application programing interface (API) functions and image processing techniques for efficient processing of satellite images. The main programing language of the course is Python. The course uses a Geospatial Data Abstraction Library (GDAL), which provides a unified way of manipulating images incorporating geospatial information. For image processing, the course uses Python-based libraries such as scikit-image.

GEOG417/GEOG617: Land Cover Characterization Using Multi-Spectral Remotely Sensed Data Sets

Teaching Assistant, Undergraduate/Graduate course, Department of Geographical Sciences, University of Maryland, 2024

This course introduces the image processing steps required for characterizing land cover extent and change. Key components of land cover characterization, including image interpretation, algorithm implementation, feature space selection, thematic output definition, and scripting are discussed and implemented. Lab exercises are designed through QGIS and Google Earth Engine.