Abstract: Land use/land cover (LULC) change monitoring is critical for understanding environmental and socioeconomic processes and to identify patterns that may affect current and future land management. Forest cover evolution in the Mediterranean region has been studied to better understand forest succession, wildfires potential, and carbon stock assessment for climate change mitigation, among other reasons. However, though multiple sources of current LULC exist, data from last century’s forest cover are less common, and are normally still reliant on locally orthophoto-interpreted data, making continuous maps of historical forest cover relatively uncommon. In this work, a pipeline based on image segmentation and random forest LULC modeling was developed to process three high resolution orthophotos (1956, 1989, and 2021) into LULC continuous land cover maps of Spain’s island of Ibiza. Next, they were combined to quantify forest evolution of Mediterranean Aleppo pine (Pinus halepensis Mill.) and to generate a continuous map of forest age classes. Our models were able to differentiate forestland with an accuracy higher than 80% in all cases, and were able to approximate forestland cover change since the mid-twentieth century, estimating 21,165 ± 252 ha (37.0 ± 0.4%) in 1956, 27,099 ± 472 ha (46.8 ± 0.8%) in 1989, and 30,195 ± 302 ha (52.8 ± 0.5%) in 2021, with a mean increase of 139 ± 6 ha (0.46 ± 0.02%, calculated from current forest cover estimate) per year. The most important variables for the identification of the forestland were the terrain slope and the image gray level or color information in all orthophotos. When combining the information from the three periods, the analysis of forest evolution revealed that a significant portion of current forest cover, approximately 15,776 ha, fell within the 75–120 year age range, while 5388 ha fell within the range of 42–74 years, and 9022 ha within the 10–41 years forest age class. Younger forests, except when mapped after known wildfires, were not considered due to the limitations of the methodology. When compared to forest age data based on ground measurements, significant differences were found among each of the remotely sensed forest age classes, with a mean difference of 13 years between the theoretical age class central value and the real observed plot average age. Overall, 63% of the forest inventory plots were assigned with the correct forest age class. This work will allow a better understanding of long-term Mediterranean forest dynamics and will help landowners and policymakers to better respond to new landscape planning challenges and achieve sustainable development goals.

Key words: Aerial orthophotos, Image segmentation, Random forest, Landscape evolution, Forest age