The study, published in the journal Discover Geoscience, was led by Rachid Ouachoua and Hamid Benssi of Ibn Tofail University in Kenitra, together with Najib Kadiri of Abdelmalek Essaâdi University in Tangier. Their approach combines multi-criteria decision-making, a structured framework for weighing competing factors, with satellite remote sensing to produce a continuous, pixel-by-pixel map of fire susceptibility across an environment where such assessments had never before been carried out. The result is not merely an academic exercise. With Morocco’s oases having lost 1,423 hectares of land and roughly 172,000 date palms to fire between 2009 and 2024, the researchers argue that their index could form the backbone of an early warning system for landscapes that are both ecologically irreplaceable and economically vital to local communities.

To construct the Fire Susceptibility Index, the team selected six conditioning factors grounded in established wildfire literature: land surface temperature, the Normalized Difference Vegetation Index, the Normalized Difference Moisture Index, proximity to roads, proximity to settlements, and land cover. Each factor was derived from satellite imagery processed at a resolution of 10 meters. Sentinel-2 surface reflectance data from 16 August 2021, ten days before the region’s largest recorded fire, supplied the vegetation indices and served as input for land cover classification, while Landsat 8 thermal imagery from 17 August 2021 was used to calculate land surface temperatures, which ranged from 32.7 to 50.5 degrees Celsius across the oasis.

A notable methodological choice was the exclusion of topographic variables such as elevation, slope, and aspect, which feature prominently in most wildfire susceptibility studies worldwide. Analysis of the shuttle radar topography data revealed that the oasis floor is almost perfectly flat, with a mean slope of just 0.59 degrees, a maximum of 5.4 degrees, and only 8 percent of the area exceeding 5 degrees of inclination. In such terrain, the researchers concluded, topography contributes essentially no spatial discrimination of fire risk, and including it would have added complexity without adding predictive power. The decision illustrates a broader principle in susceptibility mapping: the relevant drivers of fire depend on the landscape in question, and a framework designed for mountain forests must be adapted before it can serve a hyper-arid valley.

The land cover layer, arguably the most important input, was produced using a U-Net convolutional neural network, a deep learning architecture originally developed for biomedical image segmentation. The model, built on a ResNet-34 backbone with pre-trained ImageNet weights, was trained on 247 digitized training polygons spanning five classes: palms and trees, cropland, settlements, bare soil, and water. After 50 epochs of training with class balancing to correct for uneven sample sizes, the network achieved 95 percent overall accuracy, with F1 scores ranging from 93.5 percent for the palms and trees class to 100 percent for water, and a Kappa coefficient of 0.9375 indicating near-perfect agreement with reference data. The classification revealed that palms and trees dominate the oasis, covering about 2,100 hectares or 52.5 percent of the area, followed by cropland at 950 hectares, settlements at 520 hectares, bare soil at 330 hectares, and water at 100 hectares.

Weighing these factors fell to the Analytic Hierarchy Process, a technique introduced by Thomas Saaty in 1980 that converts expert judgment into numerical weights through pairwise comparisons on a one-to-nine scale. The resulting matrix, with a consistency ratio of 0.09, just under the accepted threshold of 0.10, assigned the highest weight of 0.30 to land cover, reflecting the primacy of fuel type in determining flammability. NDVI received 0.25, capturing the dense woody biomass and accumulated dry palm fronds that constitute the primary fuel in Moroccan oases. Together, fuel-related factors accounted for 55 percent of the total index weight. Proximity to roads and settlements jointly carried 31 percent, an acknowledgment that roughly 90 percent of wildfire ignitions in populated arid regions are human-caused. Moisture and temperature received smaller weights of 0.08 and 0.06 respectively, because in a landscape where summer temperatures routinely exceed 40 degrees Celsius and vegetation is consistently dry, these variables vary little and discriminate less between risky and safe locations.

Combining the weighted layers through weighted linear combination produced FSI values ranging from 0.30 to 0.91, with a mean of 0.67 and a standard deviation of only 0.07. That narrow spread is itself telling: the entire oasis system carries an elevated baseline of fire risk, a consequence of continuous dense vegetation, pervasive heat, and human infrastructure threaded through every part of the valley. Using a threshold of 0.7, selected by comparing candidate values against the known extent of the August 2021 fire, the researchers classified 42.7 percent of the oasis, approximately 1,707 hectares, as high or very high susceptibility. The spatial pattern was unmistakable, with risk climbing from the river margins toward settlements, wherever dense palm groves pressed close against roads and villages.

The validation against the August 2021 fire, which consumed 318 hectares, is the study’s most striking result. Of 28,963 burned pixels identified through the Burned Area Index calculated from post-fire Sentinel-2 imagery, 96.5 percent fell within zones the model had classified as high or very high susceptibility. Mean FSI was significantly higher for burned pixels, 0.72, than for the unburned pixels inside the same fire perimeter, 0.67, a difference that was statistically significant, though the authors caution that spatial autocorrelation between adjacent pixels means the effective sample size is smaller than the raw pixel count suggests. Receiver operating characteristic analysis yielded an area under the curve of 0.703, a figure comparable to other multi-criteria decision-making studies in Mediterranean and arid environments, and the researchers are careful to note that a susceptibility map identifies where fire conditions favor burning, not where ignition will inevitably occur.

Independent corroboration came from 60 historical fire records spanning 2009 to 2024. The mean FSI at fire points was 0.72, and no fire ever occurred in a low-risk zone. Forty-eight percent of events fell in the extreme class, 30 percent in the very high class, and 20 percent in the high class, meaning 98 percent of all recorded fires occurred in high, very high, or extreme susceptibility zones. The highest concentration of events was observed in the Aoufous sector, which recorded three fires per square kilometer over the 15-year period, confirming it as the valley’s most critical fire hotspot. This spatial agreement, the researchers argue, demonstrates that the fire risk structure identified by the index is robust and persistent, dominated as it is by relatively stable factors such as land cover and proximity to human infrastructure.

The practical implications extend well beyond the map itself. The researchers recommend systematic removal of dry palm fronds and agricultural residues, which accumulated after the region’s transition from traditional wood fuel to bottled gas, along with green firebreaks around settlements, strategic vegetation reduction in extreme-risk areas, public awareness campaigns, regulation of agricultural burning, and improved surveillance along road corridors. They also emphasize that the August 2021 fire followed extended high temperatures, underscoring the need to integrate real-time weather data with the static susceptibility map. The team acknowledges limitations, including reliance on a single large validation fire, the absence of dynamic meteorological variables such as wind speed and humidity, and the potential subjectivity of a single-author pairwise comparison matrix. Future work should incorporate multi-expert weight elicitation, higher-resolution thermal data, and additional burn indices. Even so, the study demonstrates that a transparent, interpretable framework can deliver actionable fire intelligence for data-sparse arid regions, and its integration into a dynamic geodatabase could support an operational early warning system for some of North Africa’s most vulnerable landscapes.

Subject of Research: Wildfire susceptibility assessment of the Lower Ziz Valley oases in Morocco using multi-criteria decision-making and remote sensing

Article Title: Integrating multi-criteria decision-making and remote sensing for wildfire susceptibility assessment of the Lower Ziz Valley oases in Morocco

Article References: Ouachoua, R., Benssi, H., & Kadiri, N. (2026). Integrating multi-criteria decision-making and remote sensing for wildfire susceptibility assessment of the Lower Ziz Valley oases in Morocco. Discover Geoscience, 4(1), Article 373. https://doi.org/10.1007/s44288-026-00735-8

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00735-8

Keywords: wildfire susceptibility, Fire Susceptibility Index, oasis ecosystems, multi-criteria decision-making, Analytic Hierarchy Process, remote sensing, GIS, U-Net deep learning, Land Surface Temperature, NDVI, Lower Ziz Valley, Morocco