Predicting biomass production of Typha spec. on a rewetted paludiculture site over time using multitemporal and multispectral UAS data

Peatlands, drained for agriculture and peat extraction, are carbon sources contributing 5% to global greenhouse gas emissions. In order to combat climate change and to meet carbon neutrality, peatlands need to be rewetted right away. Sustainable land-use alternatives such as paludiculture are needed for these rewetted sites. Paludiculture enables the production of biomass on rewetted peatlands while lowering emissions and further enhancing ecosystem services. Still, the applicability of paludiculture needs to be investigated in pilot schemes. To track the effectiveness of rewetting and crop growth, monitoring concepts are required. Data from Unmanned Aerial Systems (UAS) can help in predicting biomass by contributing information on spatial patterns of crop growth while keeping the workload of harvesting samples realistic. Here, the ability of optical UAS systems to provide both, spectral and structural information appears especially promising.

On a test site of 8 ha in Mecklenburg-Western Pomerania, Germany, a pilot scheme was established in 2019 (Paludi-PRIMA project) using Typha latifolia and T. angustifolia as target species. We monitored biomass production of Typha sp. using multispectral imagery (Blue, Green, Red, Red Edge, Near Infrared) and a Digital Surface Model (DSM), obtained from the UAS data with structure for motion.

We predicted Typha biomass for three different months (Jul, Aug, Sep), to evaluate the influence of phenology on prediction accuracy. In order to make best use of the different data properties, we combined a Typha mask from the multispectral imagery from all three dates (Random Forest classification with 82.5% overall accuracy and above) with structural information from the DSM.

Biomass was predicted by regression models using training data from in-situ harvests of Typha in 1-m² square plots. For these plots, spatial metrics were derived for selected UAS data derivates, e.g. the median of vegetation height from the DSM or of the NDVI. The resulting regression models were then applied to rasters representing the same metrics for the full study area. Results were validated using R², RMSE and MAE and reference information that was independently predicted from field measurements (height and shoots) for the respective observation dates.

Biomass prediction worked best with the DSM max throughout the months, with highest accuracies in August (R²=0.68 and above, RMSE<150 g/m²). The application of the Typha mask improved results for all regression models, not only for Typha-free but also surfaces with mixed vegetation cover.

We conclude that UAS data contributes essentially to biomass monitoring on experimental paludiculture sites. The combination of structural and spectral information, e.g. in the form of structural metrics and a spectral-based species mask, uses the advantages of UAS data. For larger areas the present findings need to be integrated with spaceborne data, e.g. hyperspectral satellites that add further information to the modelling.