Soil organic carbon estimation using spaceborne hyperspectral composites on a large scale
Soil Organic Carbon (SOC) is a key property for soil health. Spectral reflectance such as multispectral and hyperspectral data could provide efficient and cost-effective retrieval of SOC content. However, constrained by the availability of hyperspectral satellite data, current works mostly use a sma...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-06-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225001517 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Soil Organic Carbon (SOC) is a key property for soil health. Spectral reflectance such as multispectral and hyperspectral data could provide efficient and cost-effective retrieval of SOC content. However, constrained by the availability of hyperspectral satellite data, current works mostly use a small number of spaceborne hyperspectral imagery for SOC retrieval on a small scale. In this work, the first large-scale hyperspectral imaging reflectance composites were built, and they were used for SOC estimation. Specifically, DESIS satellite images were used to predict SOC over the whole state of Bavaria in Germany (∼ 70,000 km2). We prepare 850 hyperspectral images from the DESIS satellite and build temporal composites from them. For the soil data, data was gathered from LfU(Bavarian State Office for the Environment), LfL(Bavarian State Research Center for Agriculture) and LUCAS 2018 (Land Use and Coverage Area Frame Survey). 828 soil samples were selected after data filtering. For this regression task, different machine learning and deep learning methods were implemented and explored. Moreover, a spectral attention mechanism was added to the model. Besides hyperspectral input, the digital elevation model (DEM) was also included as an auxiliary input as the measured spectrum has inter-variability dependent on the elevation and the generated topographical features are also relevant with SOC distribution. Based on the regression results evaluated by RMSE, R2, and RPIQ, the deep learning models showed much better performance than machine learning methods. Especially when only using hyperspectral data as input, the best result was achieved with RMSE 1.947%, R2 0.626, and RPIQ 1.710 on the test set. After incorporating topographical features, the fused model achieved further improved performance with RMSE 1.752% and R2 0.695 and RPIQ 1.919. From the interpretability analysis for model performance, it was found out that the bands in the range of 530 nm–570 nm, 770 nm–790 nm, and 840 nm - 870 nm are the most relevant bands for SOC estimation. In the end, several SOC maps were generated and analyzed together with soil types. The SOC maps indicate that water-associated areas, such as coastal soils and bogs, tend to have higher SOC, while mountain areas tend to contain lower SOC. Such findings align with SOC distribution across soil types and show the effectiveness of the model. |
|---|---|
| ISSN: | 1569-8432 |