A transition in the theoretical and methodological paradigms for the rehabilitation of hardened pyroclastic formations in tropical Andean landscapes

Introduction Global ecological restoration commitments face political, economic, and technical challenges that limit their implementation and long‐term success. Among the technical constraints, the lack of effective management and monitoring plans highlights the need for tools that support efficient and accurate spatial management of restoration landscapes. Objectives This study provides theoretical and methodological contributions to digital soil mapping (DSM) for the classification and local prediction of suitable microsites for vegetation establishment, with the aim of improving restoration efficiency in hardened pyroclastic formations ( cangahua ) of the Tropical Andes. Methods Unmanned aerial vehicles equipped with multispectral cameras were used to collect environmental data in a mountainous peri‐urban landscape dominated by cangahua . In a ravine sub‐landscape, 350 soil samples were collected at a depth of 0.15 m. Soil organic carbon (SOC) was analyzed and predictively modeled using two machine‐learning algorithms: Random Forest (RF) and Convolutional Neural Networks (CNNs). Results SOC exhibited substantial spatial variability, with minimum, mean, and maximum values of 0.69%, 2.48 ± 0.71%, and 4.33%, respectively. In DSM predictions of SOC, the CNN model showed better performance ( r 2 = 34.81%) than the RF model (out‐of‐bag error‐based coefficient of determination, r 2 OOB = 24%). The most influential variables in the CNN model were the red, near‐infrared, and blue spectral bands. Conclusions DSM proved effective for identifying spatial patterns of SOC and suitable restoration microsites in cangahua landscapes. At local scales, this approach provides valuable information on landscape structure, micro‐topography, and ecological and historical memory, improving understanding of how these factors interact to influence soil properties.