This study explores a hybrid AI framework for streamflow forecasting that integrates physically based hydrological modeling, bias correction, and deep learning. HEC-HMS simulations generate synthetic discharge, which a machine learning-based bias correction model adjusts for irrigation-induced discrepancies—improving the Nash–Sutcliffe Efficiency (NSE) from 0.55 to 0.84, the Kling–Gupta Efficiency (KGE) from 0.67 to 0.89, and reducing the RMSE from 1.084 to 0.301 m³/s. The corrected discharge is used as input to a Temporal Fusion Transformer (TFT) trained on hourly meteorological data to predict streamflow at 24-, 48-, and 72-h horizons. In a semi-arid, irrigated basin in Northern Greece, the TFT achieves NSEs of 0.84, 0.78, and 0.71 and RMSEs of 0.301, 0.743, and 0.980 m³/s, respectively. Probabilistic forecasts deliver uncertainty bounds with coverage near nominal levels. In addition, the model’s built-in interpretability reveals temporal and meteorological influences—such as precipitation—that enhance predictive performance. This framework demonstrates the synergistic benefits of combining physically based modeling with state-of-the-art deep learning to support robust, multi-horizon forecasts in irrigation-influenced, data-scarce environments.

In heterogeneous agricultural landscapes, hyperspectral pixels often contain mixtures of soil and vegetation, making reliable sub-pixel characterization essential. This letter introduces a multi-sensor validation methodology for soil and vegetation fraction estimates using Planet’s Tanager-1 hyperspectral imagery (30 m), very-high-resolution UAV imagery, and field-measured endmember spectra. Validation combines: (i) distribution-level agreement with UAV-derived vegetation cover, (ii) spectral consistency with field reference spectra, and (iii) spatial coherence with landscape structure. We demonstrate our approach on a heterogeneous vineyard scene using an autoencoder-based spectral unmixing model (AE-SU) and fully constrained least squares (FCLS) as a standard supervised baseline. Results show that distribution-level metrics remain informative under cross-scale resolution mismatch, with AE-SU achieving closer agreement with the UAV-derived vegetation cover than FCLS (Overlap 0.80 vs. 0.72, JS 0.059 vs. 0.079). High-abundance pixels are also spectrally consistent with field-measured endmembers, while the resulting fraction maps preserve coherent spatial structure. These findings support practical cross-scale validation of unmixing outputs in mixed-pixel agricultural landscapes.

The forthcoming Copernicus Sentinel Expansion missions CHIME and LSTM offer new opportunities for soil monitoring through the complementary use of hyperspectral reflectance and thermal emissivity. This study evaluated the predictive value of visible to shortwave infrared (VSWIR; 350–2500 nm) reflectance and longwave infrared (LWIR) emissivity, used separately and in combination, for estimating soil organic carbon (SOC), clay, silt, and sand. High-resolution laboratory VSWIR reflectance and LWIR emissivity spectra from 787 soil samples across five Mediterranean pilot sites were resampled using preliminary CHIME and LSTM instrument spectral response functions. Random forest regression was evaluated using VSWIR-only and LWIR-only baselines together with three spectral-fusion strategies: early fusion, unweighted late fusion, and stacked fusion. The robustness of fusion gains was further assessed using 100 repeated site- and target-stratified 80/20 holdouts.VSWIR provided the dominant predictive information for all four properties, whereas LWIR-only models were consistently weaker. However, adding LWIR information improved prediction when integrated through suitable fusion strategies. On the fixed Kennard–Stone test set, the strongest fused models reached R^2 values of 0.80 for SOC, 0.85 for clay, 0.86 for silt, and 0.87 for sand, with the largest improvements relative to the VSWIR-only baseline observed for silt and sand. Repeated-holdout analysis confirmed that early-fusion gains were highly consistent across alternative data partitions, while stacked fusion provided additional benefits for selected properties. These results demonstrate complementary predictive information in the VSWIR and LWIR domains under controlled laboratory and mission-resampled conditions and provide a basis for future evaluation of joint CHIME–LSTM soil property retrieval.

Accurate characterization of soil properties is fundamental for quantifying terrestrial carbon stocks, land–atmosphere interactions, and assessing agroecosystem functioning and ecosystem services. Launched in August 2024, Tanager-1 is the first satellite in Planet’s hyperspectral constellation, delivering over 400+ contiguous spectral bands across the 400–2,500 nm range at a spatial resolution of 30 m. Such data is vital for monitoring vegetation and soil health. However, retrieving soil properties from satellite data in diverse agricultural landscapes remains challenging in heterogeneous croplands where soil, vegetation, and moisture vary strongly within and between fields. This is a particular issue in perennial cropping systems, such as vineyards. This research, conducted as part of the Horizon Europe AI4SoilHealth project, explores the potential of Tanager data to derive key soil properties, such as soil organic carbon (SOC) and soil texture, while addressing the confounding effects of different scenes' radiometric signatures.The methodology leverages the European Soil Data Center's (ESDAC) Land Use/Cover Area Frame Statistical Survey (LUCAS), which includes around 20,000 soil samples, as a foundation for retrieving soil properties from Tanager surface reflectance, using machine learning approaches such as partial least squares regression (PLSR). The study investigates strategies for decoupling the soil and vegetation components. For example, drawing on the PROSAIL radiative transfer model to explicitly simulate and account for the vegetation contribution in the training dataset and using an autoencoder-based spectral unmixing model to mitigate vegetation effects. and estimate bare soil reflectance from Tanager observations in a pre-processing step. Validation is supported by field-based spectrometers and laboratory analysis of physical soil samples to provide the ground truth for distinct scene endmembers.Preliminary findings suggest that the high signal-to-noise ratio of the Tanager hyperspectral sensor, when combined with artificial intelligence models, shows promising improvements in soil property retrieval accuracy.