Soil moisture effects: integrating physically based models and machine learning for enhanced retrieval and dryification strategies

Soil moisture, despite its crucial role in various agricultural processes, acts as noise in retrieving properties such as texture and soil organic carbon through spaceborne hyperspectral data. High spatiotemporal variability in moisture reduces the capability of soil monitoring. Soil moisture determines a reduction of the reflectance over the entire spectrum, which is not linear and its magnitude varies depending on the spectral region and the soil type. Within the framework of TEHRA project (an Italian Space Agency research initiative), a study was carried out to explore the combined use of MARMIT-2, a multilayer radiative transfer model of soil reflectance to estimate soil water content, and Machine Learning methods to address this challenge. Two local soil spectral libraries (SSLs), including both dry/wet samples and SMC (soil moisture content) values, collected over different locations in Italy between 2021 and 2022 (Maccarese-Pignola-Castelluccio and Jolanda di Savoia, respectively), have been used to investigate two different approaches. The first one is devoted to the retrieval of soil moisture content. By performing the inversion and the calibration of MARMIT-2 it is possible to increase the dataset by adding further wet spectra (and SMC values) for each sample of the original spectral library. The wet soil reflectance is expressed in terms of dry soil reflectance and three free parameters: the thickness of the water layer L, the surface fraction of the wet soil ε, and the volume fraction of soil particles in the water layer δ. Given a dry sample and the corresponding wet measurement, the Nelder-Mead algorithm is used to minimize a cost function.  The calibration, instead, is performed by fitting a sigmoid function following the soil-by-soil approach. The dataset is generated by varying (L, ε, δ) to simulate wet reflectances and the corresponding SMC is calculated using the sigmoid and the parameters found during the calibration. A Machine Learning Regression Model (a Multilayer Perceptron) has been trained using Maccarese-Pignola-Castelluccio plus additional libraries made available by authors of MARMIT and tested using Jolanda di Savoia. Results are very promising: MAE: 5.088; R2 score: 0.844; RMSE: 6.165. The model has been applied also to different PRISMA images proving to be coherent with respect to the values measured in laboratory included in the SSL. The second approach is to train a deep convolutional autoencoder capable of extracting the corresponding dry spectrum from a wet one. The dataset is composed by couples of wet and dry reflectances, resampled to PRISMA bands configuration and cleansed of water vapor absorption bands. The autoencoder consists of several blocks of convolutional layers, batch-normalization, and ReLU-activation functions. The downsampling is performed by average pooling and the upsampling with inverse convolutions. The model has been trained on Maccarese-Pignola-Castelluccio SSL, with additional samples added thanks to the inversion of MARMIT-2. Jolanda SSL has been kept aside to be used for testing the model (MAE: 0.04469, MSE: 0.00317, CS: 0.98). The autoencoder has been applied also to PRISMA images; however, further developments need to be carried out given the remarkable difference between simulated and real spaceborne data.