Conference Agenda

Urban trees play an important role in improving city liveability by reducing heat, air pollution, flood risk, and supporting a balanced, sustainable microclimate. Thus, detecting and monitoring urban trees are vital for an effective city management and environmental conservation. Traditional remote sensing methods rely on imagery from optical sensors but they face limitations in capturing inner tree structural information and LiDAR (Light Detection And Ranging) data can be a suitable alternative. Although point-cloud based approaches explore directly the three-dimensional information inherent in raw LiDAR data, the effectiveness of 3D descriptors and intensity values for tree detection can be assessed more thoroughly – specifically in the context of heterogenous and mixed trees compositions commonly found in urban environments. This work introduces an automatic and unsupervised approach for urban tree detection from airborne LiDAR data, combining intensity information with the omnivariance, a 3D descriptor based on eigenvalues. A two-step K-means clustering method is applied – first to identify potential tree points using intensity, then to detect actual trees using geometric feature – followed by morphological guided filtering to reduce misclassification. The testing was carried out on six different areas selected in datasets from Brazil and New Zealand. The evaluation was based on manually labelled reference data. The obtained results reveal an overall accuracy of 89% and low omission errors (6%), indicating method’s robustness across varied urban scenarios.

This paper aims to describe the rationale of forest Above Ground Biomass (AGB) and related Aboveground Carbon Stock (ACS) products developed with satellite remote sensing. Optical and radar imagery were synergistically combined to (1) build a classification map of vegetation species; (2) build a regression of cross-polarized, angle-normalized radar backscattering coefficient (γ0HV) with ground surveys for each class; (3) apply the corresponding regressions to the entire vegetation class map to retrieve AGB and ACS. The scope of these products is focused on two of Argentina's most densely forested provinces: Misiones in the northwestermost part and Chubut in the south. The classes involve native (tropical and temperate) and implanted forests and the ground survey sampled homogeneous stands of each, totalizing 193 data measurements of AGB. ACS was derived from AGB using published dasometric relations on a species-basis. The output of each product consists of five layers, three of which define the lower and upper limits of an 80% confidence level interval and its median value.

The other two are ancillary layers. The products are freely available online at the National Commission on Space Activities (CONAE) website [1].

Knowledge about the production of a forest area is essential for planning management activities, and machine learning techniques applied to remote sensing data have contributed to obtaining indirect production estimates at a reduced cost. The objective of this work was to evaluate the use of the Random Forest Regressor (RFR) algorithm in estimating the volume of Pinus taeda L. by analyzing hyperspectral data from the EnMAP orbital sensor. The model adjustment results were considered satisfactory on the test data, where the coefficient of determination (R2) was 0.586, the root mean squared error (RMSE) was 69.77 m3.ha-1 and the mean squared error (MAE) was 53.75 m3/ha. The spectral band groups that best explain the model are located within the near-infrared (NIR: ~734.65 nm to ~778.56 nm) and shortwave-infrared (SWIR: ~1780.22 nm to 1986.45 nm) regions. Finally, a production volume prediction map (m3/ha) was generated and compared with the average volume (m3/ha) of the inventory. There was no statistically significant difference between the inventory data and the volume data predicted by the map, confirming that the RFR has potential application in estimating the volume of Pinus taeda L. in southern Brazil.

Araucaria angustifolia (Bertol.) Kuntze, an iconic and endemic species of the Mixed Ombrophilous Forest, plays a key ecological and economic role within the Atlantic Rainforest. However, it is currently threatened by historical overexploitation and its sensitivity to climate change. This study examines the application of deep learning for the automated detection of A. angustifolia individuals in high-resolution imagery collected by Unmanned Aerial Vehicle (UAV) at selected sites in Santa Catarina, Brazil. The YOLOv11x model, a state-of-the-art convolutional neural network (CNN) architecture, was trained using two distinct datasets: a heterogeneous set and a more homogeneous one, the latter evaluated with K-Fold cross-validation. Results showed that model performance improved with increased data uniformity, with average precision (AP) rising from 21% to 27% and the F1-score from 54% to 61%. While detection accuracy remains below optimal levels, the findings highlight the model's potential for species identification. Enhancements in annotation quality, dataset diversity, and hyperparameter optimization are recommended to improve performance further and support more robust monitoring and conservation efforts for A. angustifolia.

High-resolution monitoring of individual plants is crucial for improving decision-making processes in precision agriculture, particularly when it comes to assessing development, nutrition, and health status. Deep Convolutional Neural Networks (DCNNs) have proven to be highly effective in classifying vegetation components from point cloud data based on geometric features. Combining radiometric information with geometric data can further improve classification accuracy. The fusion of LiDAR and spectral data has proved effectiveness for detailed plant discrimination. However, some challenges remain in fusing terrestrial LiDAR and multispectral data, with few studies focusing exclusively on ground-based sensor integration for plant-level classification. In this study, we propose using a Multi-layer Perceptron (MLP) architecture to classify terrestrial multispectral and LiDAR point cloud data collected around an apple tree. The model was trained and tested using datasets obtained via geometric alignment between multispectral images and LiDAR point clouds. Despite using a lightweight architecture with over 98% fewer parameters compared with architectures described in the literature, our approach achieved accuracies above 94%, comparable to state-of-the-art methods, and outperforming orbital fusion techniques. Among the spectral bands evaluated, the combination of image bands near 490 and 735 nm showed the best balance between accuracy and generalisation, consistently discriminating between leaf, wood, and fruit classes with over 90% of accuracy. These results demonstrate the potential of combining terrestrial data fusion with efficient MLP models for achieving precise plant-level classification in precision agriculture.

Detection and monitoring of land use and land

cover changes, obtained through accurate remote sensing data,

are of vital importance due to the impact of climate change and

the need to implement the principles of sustainable development

in modern society. In this context, this paper aimed to evaluate

the capacity of SAR data (Sentinel-1) for hierarchical

classification in heterogeneous landscapes. The methodology

consisted of the steps of SAR vegetation indices calculation,

dimensionality reduction of variables analysis, reduction of

sample noise, and classification process based on satellite image

time series and deep learning. The results demonstrated a

decrease in classification accuracy when considering scenarios

with more specific thematic classes, starting from three

hierarchical levels (C1 Scenario) to seven hierarchical levels (C3

Scenario), achieving global accuracy values of 0.975, 0.928, and

0.902, respectively. In addition, the similar volumetric

backscattering behavior of different vegetative canopies in SAR

data was a key parameter for the results obtained in this work.