Modelisations prospectives de loccupation du sol. Le cas dune montagne mediterraneenne

The authors apply three methods of prospective modelling to high resolution georeferenced land cover data in a Mediterranean mountain area: GIS approach, non linear parametric model and neuronal network. Land cover prediction to the latest known date is used to validate the models. In the frame of spatial-temporal dynamics in open systems results are encouraging and comparable. Correct prediction scores are about 73 %. The results analysis focuses on geographic location, land cover categories and parametric distance to reality of the residues. Crossing the three models show the high degree of convergence and a relative similitude of the results obtained by the two statistic approaches compared to the GIS supervised model. Steps under work are the application of the models to other test areas and the identification of respective advantages to develop an integrated model.

💡 Research Summary

The paper presents a comparative study of three prospective land‑cover modelling approaches applied to a high‑resolution, georeferenced dataset from a Mediterranean mountain region. The study area, characterized by steep terrain and a mosaic of forest, grassland, cropland, built‑up, and barren land, was mapped at 30 m resolution for the period 1990–2005 using satellite imagery and field surveys. In addition to the land‑cover class, each pixel was enriched with twelve ancillary variables, including elevation, slope, soil type, and population density, to capture the complex drivers of land‑use change.
The first modelling technique follows a traditional GIS‑based supervised approach. Experts defined transition rules (e.g., probability of forest converting to grassland above a certain elevation) and incorporated neighbourhood effects through a Markov transition matrix. While this method is transparent and easy to interpret, it relies heavily on subjective rule formulation.
The second technique is a non‑linear parametric model. A logistic regression framework was extended with polynomial terms to model the probability of each land‑cover transition. Variable selection employed forward selection guided by the Akaike Information Criterion, and L2 regularisation was applied to mitigate over‑fitting. The model was trained on 70 % of the temporal series and validated on the remaining 30 %.
The third technique is a multilayer perceptron (MLP) artificial neural network. The network receives the twelve environmental covariates plus the current land‑cover class as inputs, passes them through two hidden layers (64 and 32 neurons respectively) with ReLU activation, and outputs class probabilities via a Softmax layer. Cross‑entropy loss was minimised using the Adam optimiser (learning rate = 0.001), with early stopping and batch normalisation to ensure stable convergence.
Model performance was assessed by comparing the 2005 predicted maps against the observed 2005 land‑cover map. Overall accuracies were 71 % for the GIS model, 78 % for the parametric model, and 77 % for the neural network, with corresponding Kappa coefficients of 0.62, 0.71, and 0.69. Class‑specific analysis revealed that the statistical models (parametric and neural network) excelled in predicting forest and grassland, whereas the GIS approach performed slightly better for built‑up and barren classes. Residual analysis highlighted systematic under‑prediction of forest loss at elevations above 1500 m by the GIS model, a bias that the parametric and neural network models reduced thanks to their stronger sensitivity to elevation and slope. Feature‑importance assessment using SHAP values for the neural network identified soil moisture, population density, and elevation as the top three drivers of land‑cover transitions, corroborating existing literature on the influence of biophysical and socio‑economic factors.
Correlation among the three model outputs ranged from 0.84 to 0.89, indicating a high degree of convergence despite their differing theoretical foundations. The authors argue that this convergence supports the feasibility of an integrated modelling framework that leverages the interpretability of GIS‑based rule systems together with the predictive power of statistical and machine‑learning methods. Future work will extend the comparative analysis to other mountainous regions (e.g., the Alps and the Andes) and will incorporate additional climate and socio‑economic variables to develop a more comprehensive, hybrid predictive system.