Freshwater fish communities face pressure from pollution, habitat degradation, overexploitation, and human activities. Traditional capture-based surveys are labor-intensive, slow, and difficult to repeat across river networks. eDNA metabarcoding provides sensitive, noninvasive detection of rare species, but observations remain spatially discontinuous. Satellite remote sensing can extend these measurements, although broadband multispectral imagery represents habitat conditions only coarsely. Hyperspectral sensors provide finer spectral fingerprints, while learned satellite embeddings compress complex environmental information. Their combined value for aquatic biodiversity prediction remains uncertain. Based on these challenges, deeper research is needed into integrating eDNA, hyperspectral observations, and deep geospatial representations for river-scale fish assessment.
On July 28, 2026, researchers from Tsinghua University, Guangzhou University, Northeast Forestry University, the Aerospace Information Research Institute of the Chinese Academy of Sciences, and the University of Chinese Academy of Sciences published (DOI: 10.34133/remotesensing) their study in Journal of Remote Sensing. The team developed a multisource framework that integrates environmental DNA, Sentinel-2 imagery, Gaofen-5A hyperspectral data, Google Satellite Embedding features, and machine learning to predict fish taxonomic richness. The approach addresses the difficulty of producing timely, accurate, and spatially continuous biodiversity assessments across extensive river systems using conventional field surveys alone.
The innovation lies in combining interpretable Gaofen-5A (GF-5A) hyperspectral indices with 64-dimensional Google Satellite Embedding dataset (GSED) features that summarize multisource environmental conditions. GF-5A predictors outperformed Sentinel-2 multispectral variables, while GSED enabled several models to exceed a coefficient of determination (R²) of 0.90. Fusion performed best, reaching an R² of 0.94 and a root-mean-square error (RMSE) of 4.60. Hyperspectral indices represented water optics and riparian vegetation, whereas embeddings captured broader landscape structure and disturbance context. Ensemble tree algorithms, especially extra trees, handled these nonlinear ecological relationships more effectively than conventional linear approaches and produced more stable spatial predictions across tested river reaches more consistently and robustly.
The researchers assembled 149 eDNA samples from field campaigns and published studies across 15 river basins in China. Fish richness was represented by operational taxonomic units (OTUs) obtained through metabarcoding. Each site was linked to 97 Sentinel-2 vegetation indices, 157 GF-5A indices, 64 embedding features, or a fused 221-feature dataset. Sentinel-2 achieved a maximum R² of 0.65 with an RMSE of 11.40. GF-5A improved these values to 0.81 and 8.53, respectively. Embeddings alone reached 0.93 and 4.95, while fusion achieved 0.94 and 4.60. Important predictors reflected pigment status, productivity, riparian vegetation, landscape configuration, and aquatic optical conditions. Spatial mapping in the Diannong, Yellow, and Jinjiang river basins showed broadly stable cross-year patterns and localized differences consistent with habitat conditions and human disturbance. Extra trees generated smoother, more reliable distributions, whereas partial least squares regression and support vector regression tended to overestimate richness in some reaches during regional spatial extrapolation overall.
Proposed researcher comment for author approval: “Combining fine spectral information with deep satellite representations allows us to move beyond isolated sampling points and characterize fish biodiversity across river landscapes. The framework could help managers identify priority reaches, target field resources, and build faster monitoring systems, although broader datasets and cross-region validation remain essential.”
Water was sampled 0.5 meters below the surface, with three 550-milliliter replicates collected per site. eDNA was retained on 0.45-micrometer membranes, amplified by polymerase chain reaction (PCR), sequenced, and processed into OTUs. Satellite data underwent radiometric, atmospheric, and geometric correction before feature extraction. SelectKBest, random forest, and extra trees ranked predictors. Seven regression algorithms were trained with a 3:1 training-to-testing split, and performance was assessed using R² and RMSE across three repeated modeling experiments to assess predictive accuracy.
The framework could support biodiversity inventories, river restoration, environmental impact assessment, freshwater ecosystem monitoring, and evaluations of land-use or climate pressures. Satellite screening may help managers identify hotspots and direct field teams toward vulnerable reaches. Wider application requires more eDNA samples, broader coverage, comparable cloud-free hyperspectral imagery, and stronger cross-basin validation. Future integration with hyperspectral light detection and ranging (LiDAR) and foundation-model embeddings could combine screening with three-dimensional habitat characterization, improving ecological interpretation, regional generalization, and practical management-oriented uncertainty assessment.
###
References
DOI
10.34133/remotesensing.1069
Original Source URL
https://doi.org/10.34133/remotesensing.1069
Funding information
This research was supported by the National Key R&D Program of China (grant number: 2024YFF1307600) and the National Natural Science Foundation of China (grant numbers: U2243222 and 42201392).
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.