Chenhao Chai

Dr. Chenhao Chai


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
19 September 2026 in Remote Sensing
Verified
Journal Article
Open access
PDF file
CJChuan Jiang
Chuan Jiang
HZHenggang Zhang
Henggang Zhang
YWYiting Wang
Yiting Wang
CCChenhao Chai
Chenhao Chai
FFFan Feng
Fan Feng
LWLeijie Wang
Leijie Wang
Accurate delineation of landslides from high-resolution optical imagery is important for hazard assessment and infrastructure safety in mountainous regions. However, automatic segmentation remains challenging because landslides vary greatly in scale, morphology, surface appearance, and boundary clarity. These difficulties are particularly evident for old landslides, whose surface characteristics may become less distinct after long-term erosion, vegetation recovery, and surface modification. To establish a consistent baseline, five representative semantic segmentation models were evaluated on the Landslide Recognition along the Sichuan–Tibet Transportation Corridor (LRSTTC) dataset under a unified experimental protocol. DeepLabV3+ showed the strongest overall baseline performance and was selected as the base architecture. We then developed a Boundary-Aware Attention-Enhanced DeepLabV3+ (BAA-DeepLabV3+) framework, in which a Convolutional Block Attention Module (CBAM) is applied after Atrous Spatial Pyramid Pooling (ASPP) for high-level feature refinement, while an auxiliary boundary branch provides additional spatial supervision during training. In the controlled ablation experiment with seed 42, BAA-DeepLabV3+ improved the Intersection over Union (IoU) from 0.3644 to 0.4229 and the Dice coefficient from 0.5341 to 0.5944, with only a negligible increase in model parameters. Across three random seeds, the mean IoU increased from 0.3599 ± 0.0275 to 0.3876 ± 0.0315. External validation on the independent Bijie landslide dataset further improved the mean IoU from 0.7295 to 0.7444, with BAA-DeepLabV3+ outperforming the baseline under all three random seeds. These results indicate that attention-based feature refinement and boundary-aware supervision can work complementarily to improve landslide segmentation while maintaining a favorable balance between segmentation performance and computational efficiency.

19 September 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
ZLzheng Lu
zheng Lu
YWYe Wang
Ye Wang
XHXiaodong Huang
Xiaodong Huang
YWYiting Wang
Yiting Wang
CCChenhao Chai
Chenhao Chai
HZHenggang Zhang
Henggang Zhang
Accurate detection of critical infrastructure in high-resolution remote sensing imagery is important for infrastructure inventory, regional monitoring, and risk assessment. However, differences in object scale, spatial density, geometric form, and background complexity make it difficult to evaluate detectors using a single aggregate metric. This study develops a diagnostic evaluation framework utilizing CI-DIOR-7, a task-oriented benchmark derived from the DIOR dataset by retaining horizontal bounding-box annotations for seven infrastructure categories: Airport, Harbor, Bridge, Dam, Train Station, Storage Tank, and Windmill. The benchmark contains 23,463 images and 44,579 instances and preserves the original training, validation, and test splits. Four representative detectors—YOLO11s, RT-DETR-L, Faster R-CNN R50-FPN v2, and RetinaNet R50-FPN v2—were evaluated under a unified COCO-style evaluation protocol, modified with a maximum detection limit of 500 (maxDets = 500) to accurately evaluate highly dense infrastructure scenes. On the official test set of 11,738 images and 33,766 instances, RT-DETR-L achieved the highest mAP@0.5:0.95 of 0.3212 and the smallest validation-to-test performance drop. YOLO11s obtained an mAP@0.5:0.95 of 0.2785 while providing the lowest inference latency and the lowest false-positive burden on zero-GT images, at 0.168 false positives per image. Class-wise and difficulty-oriented analyses showed that Bridge was the most challenging category, with a mean class AP of 0.1514 and small-bridge recall ranging from 0.105 to 0.161 across the four models. Storage-tank detection was strongest in moderately dense scenes but deteriorated when more than 20 instances occurred in one image. These results show that no single architecture simultaneously achieved the highest detection accuracy, the lowest measured inference latency, and the lowest false-positive burden under the zero-GT protocol. CI-DIOR-7 therefore provides a reproducible benchmark and diagnostic framework for evaluating critical infrastructure detectors beyond overall mAP.

9 November 2026 in Remote Sensing
Verified
Journal Article
Open access
PDF file
JZJingyao Zheng
Jingyao Zheng
QCQian Cui
Qian Cui
QYQinyu Ye
Qinyu Ye
SQSongbo Qu
Songbo Qu
NLNing Li
Ning Li
Tianjie ZhaoTianjie Zhao
Tianjie Zhao
Surface soil moisture (SSM) regulates land–atmosphere water and energy exchanges over the Tibetan Plateau (TP); yet, passive microwave SSM products, such as those from the Soil Moisture Active Passive (SMAP) mission, are limited to coarse spatial resolutions of tens of kilometers. Although various downscaling approaches have been developed using radar backscatter, thermal infrared observations, and optical–thermal indicators, the value of shortwave infrared (SWIR) reflectance—whose moisture-dependent absorption characteristics offer fine-scale soil moisture information—remains insufficiently assessed across contrasting environments. To address this gap, we developed a 1 km SSM product for 2017–2020 by downscaling 36 km MCCA-SMAP estimates over the TP, combining the normalized shortwave difference soil moisture index (NSDSI) derived from MODIS reflectance with a first-order Taylor expansion and daily sensitivity calibration. The product was evaluated against ground observations from five monitoring networks spanning diverse climatic and land surface conditions using spatial, temporal, regional-scale, and spatiotemporal comparison frameworks, and was benchmarked against three published 1 km downscaled products based on radar backscatter, diurnal temperature amplitude, and soil evaporation efficiency, respectively. Compared with the original MCCA-SMAP SSM, the spatial regression slopes of the NSDSI-based downscaled product moved closer to unity at all five networks, with spatial correlations increasing at four networks particularly at SMN-WDL (from −0.04 to 0.59) and Shiquanhe (from 0.12 to 0.44), although the near-unity slopes at Shiquanhe were sensitive to unusually high estimates at SQ11. The NSDSI-based downscaled product showed lower mean station-level temporal correlations across the five networks, whereas regional-scale temporal correlations remained nearly unchanged. Relative to the existing 1 km products, the NSDSI-based product had spatial slopes closer to unity at all networks except Maqu, although correlation and error rankings varied among networks. At Maqu, where the vegetation cover is relatively high, the NSDSI-based downscaled SSM had spatial correlations below 0.2 and exhibited lower correlations and a higher ubRMSD than the radar- and thermal inertia-based downscaled products. The resulting downscaled SSM maps preserved the broad west-to-east moisture gradient and showed finer spatial detail. These multi-network results identify improved spatial pattern representation as the principal benefit of the evaluated SWIR-assisted downscaling approach, while demonstrating that gains in spatial agreement do not necessarily improve temporal performance or reduce ubRMSD.