Traditional methods for individual tree-crown (ITC) detection (image classification, segmentation, template matching, etc.) applied to very high-resolution remote sensing imagery have been shown to struggle in disparate landscape types or image resolutions due to scale problems and information complexity. Deep learning promised to overcome these shortcomings due to its superior performance and versatility, proven with reported detection rates of ~90%. However, such models still find their limits in transferability across study areas, because of different tree conditions (e.g., isolated trees vs. compact forests) and/or resolutions of the input data. This study introduces a highly replicable deep learning ensemble design for ITC detection and species classification based on the established single shot detector (SSD) model. The ensemble model design is based on varying the input data for the SSD models, coupled with a voting strategy for the output predictions. Very high-resolution unmanned aerial vehicles (UAV), aerial remote sensing imagery and elevation data are used in different combinations to test the performance of the ensemble models in three study sites with highly contrasting spatial patterns. The results show that ensemble models perform better than any single SSD model, regardless of the local tree conditions or image resolution. The detection performance and the accuracy rates improved by 3–18% with only as few as two participant single models, regardless of the study site. However, when more than two models were included, the performance of the ensemble models only improved slightly and even dropped.

Detection and classification of crown trees properties and assessment of their health status have raised much interest for the scientists from the forest and environmental sciences due to their essential role in landscape ecology and forestry management. The present study proposes a method based on consumer-based UAVs and Deep Learning techniques for detecting individual orchard trees and assessing key properties characterising their health status. In the proposed scheme, the Mask R-CNN model is used for detecting and mapping each individual tree morphometrical property such as the height and the crown width. Tree's health assessment across is based on the use of vegetation indices such as the Visual Atmospheric Resistance Index (VARI) and Green Leaf Index (GLI), computed from the visible spectrum camera mounted on the UAV platform. The use of the proposed approach is demonstrated at five different orchard tree species, namely plum, apricot, walnut, olive, and almond, located in Romania and Greece, computing a series of statistical metrics. Results returned outstanding ability of the algorithm's performance to map the individual trees and assess their health for four out of the five tree species (plum, walnut, apricot, almond) and satisfactory results for the fifth (olive trees). Overall, the study findings highlighted the promising potential of the proposed methodological framework and its scalable potential for wider applicability as a low-cost, effective solution in mapping individual trees properties and health conditions in the field.

Computer modeling of sporadic and isolated patches of mountain permafrost distribution is difficult to implement without overestimating it. The main challenge is to determine the very areas where the criteria for permafrost maintenance are met. This paper aims to modeling the permafrost distribution in the Southern Carpathians (SC), a typical marginal periglacial mountain range. For this purpose, a collection of 883 bottom temperature of late winter snow cover (BTS) points was used as a proxy for permafrost presence or absence in order to train several machine learning models. The performances of each model were evaluated with AUC with varying between 0.99 for Maxent and 0.74 for K‐nearest neighbors and most models (five) exhibiting values between 0.82 and 0.86. Other tests such as confusion matrices, sensitivity analyses, data shuffling, and data size reduction tests indicated that Maxent, AdaBoost, and support vector machine offered the best results while logistic regression, neural network, and gradient boosting exhibited rather poor permafrost distributions. The final ensemble median model indicated a total permafrost area of 19.2 km ² occupying 1%–9% of the alpine area of the studied massifs. NDVI proved crucial for permafrost prediction because it allows delimiting the debris surfaces where permafrost is probable.