Enhancing Soil Fertility Mapping with Hyperspectral Remote Sensing and Advanced AI: A Comparative Study of Dimensionality Reduction Techniques in Morocco

As global food demand increases, farming systems experience heightened pressure to enhance productivity on limited arable land. In Africa, including Morocco, smallholder farms are particularly susceptible to climate variability, soil degradation, and suboptimal farming practices, resulting in yield gaps—the disparity between actual and potential yields under optimal conditions. In Morocco, yield variability is significantly influenced by soil fertility, irrigation, and climate. Consequently, quantitative assessment and mapping of key soil fertility indicators at the field scale are essential for improving yields. Remote sensing data, particularly hyperspectral imagery, presents a cost-effective and time-efficient alternative to traditional soil mapping methods. However, its potential for providing detailed local-scale soil information in Africa remains underexplored. This study utilizes high-resolution PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral imagery and laboratory-analyzed soil samples to map four key soil properties—cation exchange capacity (CEC), soil organic matter (SOM), available phosphorus (P₂O₅), and exchangeable potassium (K₂O)—in the Ain el Orma agricultural area on the Saïss plateau, Morocco. Despite the advantages of hyperspectral sensors, their high processing complexity, due to redundant or correlated spectral bands, can impede machine learning model accuracy. This study compares the performance of traditional and advanced machine learning algorithms combined with dimensionality reduction techniques—PCA, UMAP, and RFE. Six well-established algorithms (XGBoost, Gradient Boosting, PLSR, SVR, and Random Forest) were evaluated as an initial step in the artificial intelligence workflow, yielding weak to moderate results. For SOM (%), the utilization of RFE resulted in the optimal performance with a substantial improvement in R² from 0.30 (PCA) to 0.36, while the Root Mean Squared Error (RMSE) decreased from 0.52 to 0.39%. Furthermore, the Ratio of Performance to Interquartile Range (RPIQ) for SOM (%) also increased from 1.58 (PCA) to 2.10. In the case of P₂O₅ (mg/kg), PCA emerged as the superior method, yielding an R² of 0.38 compared to 0.37 for RFE and -0.01 for UMAP. The RMSE decreased from 11.92 (RFE) to 11.82. For K₂O (mg/kg), PCA again proved to be the optimal method, with an R² improving to 0.13 from -0.29 with RFE and remaining superior to UMAP's 0.19. The RMSE decreased from 107.33 (RFE) to 88.51%, and the RPIQ increased from 1.50 to 1.82. Lastly, for CEC (meq/100g), PCA delivered the most accurate predictions, improving the R² to 0.68 from 0.60 (RFE) and 0.21 (UMAP). The RMSE was reduced significantly from 2.08 (RFE) to 1.88%, while the RPIQ increased from 2.47 to 2.73. These initial findings underscore the importance of feature selection and dimensionality reduction for developing robust models for soil property estimation using hyperspectral data. 
Additionally, this study aims to propose advanced innovative AI models capable of enhancing the accuracy of soil maps. In conclusion, the anticipated results are expected to support the creation of accurate soil maps, necessary for spatialized analysis of wheat yield variability using hyperspectral remote sensing imagery, thus contributing to food security and sustainable agricultural practices.