Markov beats neural networks in forecasting Nigeria’s fast-urbanising landscape

A comparison of two widely used land-change simulation models in Southwestern Nigeria has found that the older, simpler approach outperforms the machine-learning alternative. Between 2002 and 2022, built-up land in the study region grew by more than 340 percent while wetlands shrank by roughly 47 percent, and when researchers asked two hybrid models to reproduce the 2022 landscape from a decade of training data, the Cellular Automata–Markov (CA-Markov) model clearly beat the Cellular Automata–Artificial Neural Network (CA-ANN) model on statistical accuracy, spatial pattern and class-level stability.

The study, published in Discover Geoscience, classified multi-temporal Landsat imagery from 2002, 2012 and 2022 covering Landsat Path 190, Row 55 — an area spanning much of Osun, Ondo and Ekiti States with smaller portions of Lagos, Ogun and Oyo. The team processed the imagery in TerrSet’s Land Change Modeller, calibrating both models on the 2002–2012 interval and validating their predictions against an independently classified 2022 map. Four geographic information system-derived drivers informed the simulations: distance to roads, distance to rivers, slope and elevation, all normalised to a 0–1 scale before entering the modelling pipeline.

The underlying satellite data demanded careful preparation. The 2002 and 2012 maps came from Landsat 7 Enhanced Thematic Mapper Plus sensors and the 2022 map from Landsat 8’s Operational Land Imager and Thermal Infrared Sensor, all acquired in the dry season between December and January to minimise cloud contamination and seasonal vegetation variability. Atmospherically corrected Level-2 surface reflectance products were used for spectral consistency, clouds and shadows were removed using the Quality Assessment bands, and the well-known Scan Line Corrector gaps in the Landsat 7 imagery were repaired with a multi-date local linear histogram matching technique following United States Geological Survey guidelines. All images were then co-registered to a common coordinate system, UTM Zone 31 N on the WGS 84 datum, with sub-pixel accuracy.

Classification itself followed a supervised maximum likelihood approach implemented in ENVI, using five broad land-cover classes aligned with a modified Level I scheme and the principles of the FAO Land Cover Classification System. Between 25 and 50 homogeneous training polygons per class, each containing a minimum of 25 pixels, were digitised on-screen at a consistent viewing scale of roughly 1:15,000, guided by false-colour composites, high-resolution Google Earth imagery of matching dates and, where available, field observations. A 3 × 3 majority filter removed isolated single-pixel misclassifications, and accuracy of the classified maps was assessed with 500 independent validation samples drawn through spatially stratified sampling from high-resolution imagery and historical topographic maps.

The documented landscape transformation was striking. Built-up area expanded from 449.8 square kilometres, or 1.42 percent of the study area, in 2002 to 1,013.7 square kilometres in 2012 — a 125.35 percent jump — and then to 1,980.3 square kilometres, or 6.27 percent, by 2022. Wetlands fell from 14,965.80 square kilometres, 47.36 percent of the area, to 7,868.22 square kilometres, a 47.43 percent decline over the two decades. Open surfaces rose 171.66 percent, an expansion the authors link to quarrying, mining, deforestation and land degradation, while vegetation declined a comparatively modest 4.55 percent and water bodies increased 69.28 percent yet remained a minor share of total cover.

Spatially, the change followed predictable corridors. Urban expansion accelerated around Ibadan, Akure and Abeokuta along major highways, confirming road proximity as the dominant driver of conversion to built-up land. Among the four explanatory variables tested with Cramér’s V statistics, distance to roads showed the strongest association with vegetation-to-built-up and open-surface-to-built-up transitions, at 0.34. Distance to river, at 0.23, best explained wetland losses to open surface and vegetation, consistent with agricultural encroachment along the Osun River basin, while slope and elevation contributed less explanatory power, at 0.14 and 0.18 respectively, reflecting the region’s generally low relief outside the Idanre Hills.

Against this backdrop, the head-to-head model comparison produced a clear verdict. Simulating 2022 conditions, CA-Markov achieved an overall accuracy of 89.4 percent and a Kappa coefficient of 0.846, while CA-ANN returned 78.9 percent overall accuracy and a Kappa of 0.707. Both figures clear the conventional threshold of 0.70 for acceptable agreement, but the disaggregated results revealed deeper problems for the neural network approach. CA-Markov projected 2,265.08 square kilometres of built-up land, overestimating the observed extent by 14.38 percent, and also over-predicted wetlands by 21.33 percent. CA-ANN, by contrast, drastically underestimated built-up area, predicting only 861.41 square kilometres — a shortfall of 56.50 percent relative to the observed map — while overestimating open surfaces by 20.17 percent and underestimating water bodies by 53.03 percent.

The error structure of each model told its own story. Disaggregating the confusion matrices into producer’s and user’s accuracies showed that CA-ANN’s failures were concentrated in omission error: pixels that were genuinely built-up, water or vegetation in the observed 2022 map were frequently allocated to other categories. In other words, the neural network failed to recover existing built-up and water pixels rather than indiscriminately over-assigning other classes to them, pointing to insufficient discriminatory power in the four available drivers within its multilayer perceptron transition function. CA-Markov, meanwhile, displayed comparatively balanced omission and commission errors across classes, which translated into its higher Kappa score and a more coherent spatial output.

Those spatial differences were visually evident in the agreement maps. CA-Markov reproduced contiguous urban growth along the major transportation corridors and more realistically captured wetland contraction zones, whereas CA-ANN produced scattered, discontinuous fragments of simulated built-up expansion and tended to preserve wetlands in areas that had in reality already been converted. The authors attribute the neural network’s weaker showing to its known sensitivity to driver selection and training data quality, noting that suboptimal calibration often results in spatially dispersed predictions. A McNemar’s chi-squared test applied to the paired confusion matrices was used to assess whether the accuracy difference between the models was statistically significant.

Methodologically, the study deliberately went beyond the single aggregate Kappa value that dominates much of the land-change literature. Reliability was defined as a composite measure combining statistical agreement, spatial consistency between simulated and observed maps, and class-level stability of land-cover categories over time. The authors argue that overall accuracy alone can mask class-specific errors and allocation disagreements, and their results support recent recommendations that land-change model evaluation explicitly report quantity and allocation disagreement alongside class-level omission and commission errors. They also note that including multiple driving variables contributed to better performance than comparable studies that omitted such drivers; one earlier study that found the two models comparable did not incorporate land-change driving forces at all, a limitation the authors suggest likely influenced its outcome.

The choice of study setting matters for the wider implications. Southwestern Nigeria, home to major urban centres such as Lagos and Ibadan and experiencing an annual population growth rate of 3.2 percent, is one of the most dynamic land-change regions in the country, yet predictive modelling there remains underexplored relative to its importance. In data-constrained African contexts, where budgetary and technical limits restrict access to high-resolution imagery and dense ancillary datasets, identifying which hybrid cellular-automata approach delivers dependable results under realistic conditions carries practical weight. The CA-Markov Kappa of 0.846 achieved here is closely comparable to Kappa values reported in other regional studies, and the authors suggest that adding more driving factors could push the result higher still.

Physically, the region adds its own complexity to any simulation exercise. It is underlain predominantly by Precambrian Basement Complex rocks with widespread lateritic soils, and its terrain rises from roughly 200 feet to more than 3,000 feet at the Idanre Hills in Ondo State. The Osun River constitutes the principal drainage system, flowing southward into the Lagos Lagoon, while the climate is bimodally rainfall-fed, with mean annual totals ranging from about 1,237 millimetres in Abeokuta to 1,507 millimetres in Lagos and mean annual temperatures around 28 degrees Celsius. Vegetation transitions from lowland rainforest in the south, including freshwater swamp forests along major rivers, to southern Guinea savanna in the north, though recurrent disturbance through agriculture, logging and burning has prevented stratified forest development and favoured secondary, grass-derived savanna.

The authors conclude that CA-Markov provided the more reliable simulation under the conditions tested, aided by its ability to prioritise a specific class — in this case built-up expansion along major roads — as the dominant driver of change. CA-ANN, they argue, requires additional explanatory variables, parameter optimisation and further calibration before it can be considered a dependable alternative in this setting. Looking ahead, they recommend incorporating socio-economic factors, policy variables and demographic trends into future models, acquiring higher-quality ground-truth data through field surveys, and fostering collaboration between remote-sensing specialists, environmental scientists, urban planners and policymakers. Under a continuation of existing development trends, the validated models were also used to generate projections for 2032, offering planners in the region a tested, transparent tool for anticipating where the next decade of change is likely to fall.

Subject of Research: Comparative performance of CA-Markov and CA-ANN models for land use and land cover change simulation in Southwestern Nigeria

Article Title: Evaluating the performance and reliability of CA-Markov and CA-ANN models for land use/land cover change simulation in Southwestern Nigeria

Article References: Akintuyi, A., Wunude, E., Fashoto, B., & Adedo, L. (2026). Evaluating the performance and reliability of CA-Markov and CA-ANN models for land use/land cover change simulation in Southwestern Nigeria. Discover Geoscience, 4(1), Article 414. https://doi.org/10.1007/s44288-026-00790-1

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00790-1

Keywords: land use land cover change, CA-Markov, CA-ANN, remote sensing, Landsat, urban expansion, wetland loss, Southwestern Nigeria, cellular automata, Kappa coefficient, Land Change Modeller, GIS