Agricultural soil properties mapping from PRISMA and EnMap data: exploiting multitemporal bare soil approaches

Challenge Understanding the properties of agricultural soils is essential for optimising food production and the overall efficiency of agricultural systems. To obtain key information on soil characteristics, remote sensing techniques are increasingly employed. By analysing the distinct "spectral signature" of materials, it is possible to accurately estimate these characteristics in a quantitative manner. The Italian PRISMA and the German EnMAP hyperspectral satellite missions are expected to greatly contribute to the improved knowledge of agricultural soil properties spatial variability and mapping. In this study, multiple bare soil PRISMA and EnMAP hyperspectral images acquired over agricultural fields in Northern Italy were used to retrieve Soil Organic Matter (SOM), CaCO3, and texture topsoil properties using Machine Learning Regression Algorithms and were validated using an adapted field measurement scheme for soil sampling and analysis. Methodology The study site is located in the Jolanda di Savoia farm (Lat. 44.87°N, Lon. 11.97°E) (Italy). Soils are characterised by significant variability, with the presence of buried paleo-channels dating back to the past presence of marshland, in which land reclamation has been carried out since the end of 1800s. Between 2019 and 2023, respectively, 15 PRISMA and 6 EnMAP images were acquired. Concurrently, topsoil samples were collected on 32 agricultural fields, according to Elementary Sampling Units (ESU) of 30 by 30 m, followed by laboratory analyses of SOM, CaCo3, texture and pH. To select only bare soil signals, excluding the presence of photosynthetic or non-photosynthetic vegetation, various processes were employed. These included the combined utilization of spectral indices such as the Normalized Difference Vegetation Index (NDVI) and the non-photosynthetic vegetation index (nCai). Two different spectral libraries were assembled, using respectively, PRISMA and EnMap bare soil reflectance acquired on different dates, associated with the corresponding soil properties. The optimal samples were selected from the soil libraries, employing several criteria, such as prioritizing those exhibiting the lowest soil moisture levels. The dataset underwent different preprocessing techniques to reduce the impact of noisy bands. Different machine learning regression algorithms were then trained on these libraries using a K-fold cross-validation approach. Results Different methodologies of spectra extraction from the different acquisition dates available for PRISMA and EnMAP were tested, along with different preprocessing approaches and machine learning algorithms, in various combinations. The results showed that selecting dates from a multitemporal soil reflectance set, based on soil moisture and applying spectral pre-treatments, can generally improve the accuracy of predictions, though the best pre-treatment varied depending on the soil properties and sensor. The multi-date approach needs further refinement as the indexes used do not consistently select the driest soil for both datasets. Partial Least Squares Regression (PLSR) was found to be a highly consistent and successful algorithm for prediction. For Clay, the best results were achieved using the PRISMA sensor, Relevance Vector Machine algorithm, and Derivative of first-order preprocessing technique, resulting in an R2 of 0.82 and a root mean square error (RME) of 6.72%. For Silt, the EnMap sensor, PLSR algorithm, and Derivative of first-order preprocessing technique yielded an R2 of 0.81 and RME of 5.49%. For SOM, the PRISMA sensor and PLSR algorithm provided the best results, with a R2 of 0.77 and RME of 1.84%. Lastly, for CaCO3, the best results were obtained by PRISMA using a Standard Normal Variate preprocessing technique applied to absorbance smoothed with Savitzky-Golay filter of the first, yielding an R2 of 0.58 and RME of 2.60%. Outlook for the future These preliminary tests have yielded promising results, as PRISMA and EnMAP time series were both able to retrieve most soil parameters with comparable accuracy, considering the smaller number of EnMAP acquisitions available than for PRISMA. Repeated bare soil acquisitions from both sensors could be valuable and synergic for agricultural soils mapping, though further testing could help identify the strengths and weaknesses of each multitemporal data extraction approach. Overall, this offers a constructive step towards improving our ability to understand soil characteristics and improve agricultural and environmental practices. Forthcoming investigations should focus on the development of more general methodologies. In this context, good opportunities arise from the use of open lab-based soil spectrum libraries (OSSL) in methodologies such as Spiking or reinforced learning, as well as the integration of data from proximate sensors through the application of data fusion techniques.