Keli Cheng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Keli Cheng has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Grant J. Scott
Institute for Data Science and Informatics, University of Missouri, Columbia, MO 65211, USA
2 shared publications - Ilinca Maria Popescu
Department of Geography, Stanford University, Stanford, CA 94305, USA
1 shared publication - Lincoln Sheets
MU Institute for Data Science and Informatics (MUIDSI), University of Missouri, Columbia, MO, USA
1 shared publication - Weipeng Wu
Department of Personal Financial Planning, University of Missouri, Columbia, USA
1 shared publication
Latest Publications
- 1 April 2026
- Published by Springer Nature
- p. 608-617
Benchmarking Few-Shot Methods for Rare Target Classification in PlanetScope Imagery
- 1 April 2026
- Published by Springer Nature
- p. 608-617
Benchmarking Few-Shot Methods for Rare Target Classification in PlanetScope Imagery
- Open Access
- 26 September 2023
- Published by MDPI AG in Remote Sensing
- Vol. 15 (19) , 4705
Deep Seasonal Network for Remote Sensing Imagery Classification of Multi-Temporal Sentinel-2 Data
As a medium-resolution multi-temporal data source, Sentinel-2 data has the potential to match the performance of using very-high-resolution (VHR) images in deep learning applications. To fully leverage the multi-temporal nature of Sentinel-2 data, we introduce the Deep Seasonal Network (DeepSN). This composite architecture combines a pre-trained deep convolutional neural network (DCNN) for visual feature extraction with a long short-term memory (LSTM) model to capture temporal information and make classification predictions. We evaluate the effectiveness of DeepSN on a Maasai Boma classification task in the Tanzania region. The DeepSN takes a sequence of four seasonal data, each spanning three months, for Boma prediction. Through cross-season validation experiments, we compare various advanced DCNNs and select EfficientNet as the backbone for DeepSN, as it performs the best. DeepSN with an EfficientNet backbone achieves a significant 19% improvement in the F1 score compared to plain EfficientNet for the Boma classification task. This work introduces a versatile composite architecture capable of handling multi-temporal data efficiently, providing flexibility in choosing the most suitable feature extraction backbone. The performance of DeepSN demonstrates the viability of utilizing medium-resolution multi-temporal data instead of high-resolution images for diverse tasks.
- 26 September 2023
- Published by MDPI AG in Remote Sensing
- Vol. 15 (19) , 4705
Open AccessDeep Seasonal Network for Remote Sensing Imagery Classification of Multi-Temporal Sentinel-2 Data
As a medium-resolution multi-temporal data source, Sentinel-2 data has the potential to match the performance of using very-high-resolution (VHR) images in deep learning applications. To fully leverage the multi-temporal nature of Sentinel-2 data, we introduce the Deep Seasonal Network (DeepSN). This composite architecture combines a pre-trained deep convolutional neural network (DCNN) for visual feature extraction with a long short-term memory (LSTM) model to capture temporal information and make classification predictions. We evaluate the effectiveness of DeepSN on a Maasai Boma classification task in the Tanzania region. The DeepSN takes a sequence of four seasonal data, each spanning three months, for Boma prediction. Through cross-season validation experiments, we compare various advanced DCNNs and select EfficientNet as the backbone for DeepSN, as it performs the best. DeepSN with an EfficientNet backbone achieves a significant 19% improvement in the F1 score compared to plain EfficientNet for the Boma classification task. This work introduces a versatile composite architecture capable of handling multi-temporal data efficiently, providing flexibility in choosing the most suitable feature extraction backbone. The performance of DeepSN demonstrates the viability of utilizing medium-resolution multi-temporal data instead of high-resolution images for diverse tasks.
- Open Access
- 13 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 15 (19391404) , 3916-3924
Analysis of Deep Learning Techniques for Maasai Boma Mapping in Tanzania
Underdeveloped countries in sub-Saharan Africa often contain cultural sub-populations that are under-served in regards to health and education. This perpetuates the health challenges of the country as a whole, and it is therefore of interest to be able to automatically map the sub-population for health services delivery. Non-profit health organizations have taken the lead in these efforts, providing humanitarian aid as well as health care services. This is necessary, as the ethic sub-populations are not well integrated into the society and the existing healthcare infrastructure and systems. In this study, we explicitly explore the Maasailand of Tanzania, to evaluate the use of deep neural networks (DNN) to aid in the automatic visual analysis of remote sensing data to geo-locate Maasai boma structures. We investigate the performance of four diverse state-of-the-art DNN as scene classifiers of boma presence within high-resolution imagery; all showing over 95% F1 score performance. Additionally, we scan over 3900 km2 of high-resolution imagery, combining a top-performing DNN, ProxylessNAS, with broad area aggregation and mapping techniques and demonstrate the discovery of hundreds of boma, many that were not discovered by human analysts performing visual scans. The trained ProxylessNAS model was used to generate a classified vector response field (CVRF). The CVRF was then aggregated by a mode-seeking algorithm to detect potential locations of boma structures within the study area. The model detected numerous human false negatives (HFNs) and achieved 94.022% TPR and 95.395% F1 score using an aggregation aperture of 250 m within a 76.620 km2 area.
- 13 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 15 (19391404) , 3916-3924
Open AccessAnalysis of Deep Learning Techniques for Maasai Boma Mapping in Tanzania
Underdeveloped countries in sub-Saharan Africa often contain cultural sub-populations that are under-served in regards to health and education. This perpetuates the health challenges of the country as a whole, and it is therefore of interest to be able to automatically map the sub-population for health services delivery. Non-profit health organizations have taken the lead in these efforts, providing humanitarian aid as well as health care services. This is necessary, as the ethic sub-populations are not well integrated into the society and the existing healthcare infrastructure and systems. In this study, we explicitly explore the Maasailand of Tanzania, to evaluate the use of deep neural networks (DNN) to aid in the automatic visual analysis of remote sensing data to geo-locate Maasai boma structures. We investigate the performance of four diverse state-of-the-art DNN as scene classifiers of boma presence within high-resolution imagery; all showing over 95% F1 score performance. Additionally, we scan over 3900 km2 of high-resolution imagery, combining a top-performing DNN, ProxylessNAS, with broad area aggregation and mapping techniques and demonstrate the discovery of hundreds of boma, many that were not discovered by human analysts performing visual scans. The trained ProxylessNAS model was used to generate a classified vector response field (CVRF). The CVRF was then aggregated by a mode-seeking algorithm to detect potential locations of boma structures within the study area. The model detected numerous human false negatives (HFNs) and achieved 94.022% TPR and 95.395% F1 score using an aggregation aperture of 250 m within a 76.620 km2 area.