Conference Agenda

This study aims to characterize the spectral reflectance of healthy and huanglongbing (HLB)-infected citrus individuals at both the leaf and plant levels using a statistical approach. Our main contribution is to assess the extent to which hyperspectral measurements can differentiate disease status. Spectral data were collected from 1,912 leaves belonging to 89 citrus plants, of which 29 were found to be infected with HLB and 60 were healthy. A statistical protocol—including Shapiro-Wilk, Welch’s t-tests, ANOVA, and Z-tests—was applied to estimate the mean and standard deviation of spectral reflectance for each class, evaluate the spectral variance across bands at the plant level, determine differences between HLB-positive and HLB-negative groups at both hierarchical levels (leaf and plant), and identify the spectral bands with the highest discriminatory power. The findings reveal substantial intra-plant spectral variability in HLB-positive citrus, indicating that individual leaf reflectance may not reliably represent whole-plant disease status. This reinforces the need for plant-level spectral aggregation in remote sensing models. Discriminative spectral intervals were consistently identified in the 400–431 nm, 488–752 nm, 1132–1830 nm, and 1890–2500 nm ranges, spanning the visible to shortwave infrared (SWIR) spectrum.

Nitrogen content is essential for grass growth, grassland management, and forage productivity. In general, the nitrogen amount is indirectly estimated using manual techniques for sample acquisition and laboratory analysis, which are a costly endeavour, mainly in large agricultural areas. In this context, remote sensing technologies allow monitoring important parameters for agriculture, fast, non-destructively, and on a large scale, using aerial images obtained by Unmanned Aerial Vehicles (UAV) and analysed through spectral indices and structural variables of the vegetation. However, further studies are needed that use more affordable sensor systems that can be used in large areas, such as in Brazil. This work asses the feasibility of employing GoPro wide-angle RGB camera onboard a UAV to estimate the nitrogen content of an experimental grassland area. Different data scenarios were evaluated, incorporating combinations of vegetation indices (VIs) and three-dimensional (3D) metrics derived from the Canopy Height Model (CHM): all available metrics (ALL), a subset of three VIs combined with four 3D metrics (VI3 + CHM4), and 3D metrics only. To estimate nitrogen content, the Random Forest (RF) machine learning algorithm was applied. The most accurate model, yielding the lowest error, resulted from integrating data from two acquisition dates, achieving a coefficient of determination (R²) of 0.83 for the model, a Pearson Correlation Coefficient (PCC) of 0.82 in the validation trials, and a Root Mean Square Error expressed as a percentage (RMSE%) of 19.62%. These findings highlight the potential of UAV-mounted RGB sensors as an effective tool for estimating pasture parameters.

Phytonematodes are parasites of plant tissues, primarily roots. Although they cause significant damage to agriculture, they are still underexplored. Soybean crops suffer significant yield losses due to phytonematode infestations. Remote sensing techniques offer an effective means of detecting and analyzing these infestations. This study aimed to identify the most suitable stage in the soybean plant's growth cycle for detecting spectral differences between nematode-infested and healthy plants using MSI/Sentinel-2 imagery. The study utilized three sets of Sentinel-2 images, representing different growth stages (December - vegetative, January - grain development, February - maturation). 38 nematological sample points were classified into three infestation levels (no or low, moderate, high). Spectral responses were collected for each growth stage, followed by an analysis of reflectance values and NDVI to assess class separability. The maturation stage exhibited the most significant spectral differences between infestation levels, especially in the near-infrared and red-edge regions. The sum of squared differences demonstrated the greatest separability in the maturation stage, particularly between moderate and no or low infestation categories. The results suggest that the end of the reproductive stage is optimal for detecting spectral differences in soybean plants affected by nematodes.

Brazil stands as the world's leading coffee producer and exporter, playing a crucial role in global supply. Considering its importance in the national agricultural scenario, this study aims to evaluate the potential of features derived from C-band Synthetic Aperture Radar (C-SAR) Sentinel-1 imagery in detecting structural variations in coffee-growing plots of different cultivars and ages. The research conducts a comparative analysis of SAR polarimetric attributes and the Normalized Difference Vegetation Index (NDVI), utilizing optical imagery from the Multispectral Instrument (MSI) aboard Sentinel-2B. The experimental area is situated within a commercial coffee plantation owned by Fazenda Juliana, near the municipality of Monte Carmelo, Minas Gerais. The farm has supplied precise delimitation of coffee plots, along with detailed records on planting age and cultivars. C-SAR Sentinel-1B were preprocessing using the Sentinel Application Platform (SNAP), generating backscatter coefficients (𝜎⁰) in VH and VV polarizations and key polarimetric indices Cross Ratio (CR), Normalization Ratio (NL), Radar Gap Index (RGI), and Radar Vegetation Index (RVI), were calculated. SAR features and NDVI were segmented by the defined coffee plot boundaries, and for each segmented plot, the mean values of 𝜎⁰ VH, 𝜎⁰ VV, SAR indices, and NDVI were extracted for analysis. The findings demonstrate a strong correlation (>0.9) between NDVI fluctuations and 𝜎⁰ VH and NL variations, underscoring their effectiveness in detecting structural differences across coffee plantations. This approach offers a robust alternative for monitoring crop variability, particularly in areas where optical sensing is constrained.

Pollinators play a central role in sustaining biodiversity and ecosystem services, consequently their response to forest regeneration in tropical landscapes needs to be quantified at large scales. Here, we assess how land cover composition and forest age influence pollinator diversity in the Brazilian Atlantic Forest — a global biodiversity hotspot undergoing extensive regeneration. We integrated land-use and forest age data from MapBiomas with 56,593 bee occurrence records from GBIF, focusing on five bee families. Using Random Forest models, we evaluated the importance of land cover types and secondary forest age intervals for predicting total occurrences and genus richness. Our results show that primary forest cover is the dominant predictor of bee genus richness, followed by late-stage secondary forests aged $>26 $ years and riparian-associated water surfaces. In contrast, younger secondary forests ($<25$ years) contributed negligibly and urban dominated landscapes support less diversity overall. While total occurrence data reflected strong spatial bias towards non-vegetated and agricultural areas, genus richness emerged as a more robust parameter, avoiding bias, and mitigating over-representation from anthropic landscapes. Our findings highlight the ecological value of mature secondary forests for pollinator conservation and reinforce the need to incorporate the time dimension into restoration monitoring. Our results underscore the conservation value of mature secondary forests and the need to integrate forest age into restoration monitoring. Our approach demonstrates the utility of combining biodiversity data, geospatial data derived from remote sensing, and machine learning to produce scalable, spatially explicit insights into ecological recovery and pollination services in tropical biomes.

Banana crop is highly susceptible to Fusarium wilt, a disease that can cause significant agricultural losses if not detected early. This study aimed to develop a classification model to detect Fusarium-infected banana plants using multispectral imagery. The dataset consisted of labeled images categorized into two classes: healthy and diseased (Fusarium). A convolutional neural network (CNN) was trained and evaluated, achieving an overall test accuracy of 81.25%. Class-wise evaluation showed a precision of 70%, recall of 100%, and F1-score of 82% for healthy plants, while the diseased class reached 100% precision, 67% recall, and 80% of F1-score. These results indicate strong performance in precision but highlight a need to improve recall for effective disease monitoring. Comparisons with recent studies show that higher accuracy can be achieved through larger datasets, data augmentation, and transfer learning. This research demonstrates the potential of using multispectral images and deep learning for banana disease detection, with future improvements focused on expanding data volume and applying advanced training techniques to boost recall and overall robustness.