https://www.alice.cnptia.embrapa.br/alice/handle/item/196 Artigo em periódico indexado (CNPTIA) 2026-08-14T00:22:51Z https://www.alice.cnptia.embrapa.br/alice/handle/doc/1189109 Título: Data-centric approach for land use and land cover classification in Brazil. Autoria: VAZ, G. J.; ANTUNES, J. F. G.; ESQUERDO, J. C. D. M.; COUTINHO, A. C.; TAVARES, A. S.; CALABONI, A.; BERTOLO, L. S.; FELIX, F. C.; ROCHA, A. Conteúdo: Abstract. Land use and land cover (LULC) classification plays a crucial role in addressing numerous real-world challenges. Hence, we proposed methodological advances in LULC classification from a data-centric artificial intelligence perspective, which prioritizes data quality as a key factor in improving machine learning performance. The main contributions include evaluations of novel approaches for: (i) constructing an accurately labeled dataset based on agreement among existing reliable maps; (ii) curating remote sensing data to improve accuracy, consistency, unbiasedness, relevance, diversity, and completeness; (iii) generating training samples that capture the spatial, temporal, and spectral dimensions of remote sensing data; and (iv) developing a deep learning model designed to leverage multidimensional features. The study evaluates a sample generation method grounded in reference map agreement and multidimensional feature extraction, along with a deep learning model that leverages these features, attaining high accuracy across all LULC classes and providing a robust basis for large-scale, data-centric LULC mapping. 2026-01-01T00:00:00Z https://www.alice.cnptia.embrapa.br/alice/handle/doc/1188987 Título: Climate adaptation and economic perspectives on anti-hail net adoption in apple orchards: an exploratory case study in southern Brazil. Autoria: FURUYA, D. E. G.; BOLFE, E. L.; SOARES, V. B.; SILVEIRA, F. da; BARBEDO, J. G. A.; GEBLER, L. Conteúdo: Hail events represent a major source of economic risk in apple production, leading to significant losses in yield and marketable output. Although anti-hail net systems are increasingly adopted as a mitigation strategy, evidence on their economic viability remains limited, especially in emerging regions. This study examines the economic dimensions of anti-hail net adoption in apple orchards in Vacaria, southern Brazil, by integrating remote sensing data, official loss records, and producer interviews. Multitemporal Sentinel-2 imagery was used to map the expansion of net-covered areas between 2016 and 2026. Economic loss data from eight recorded hail events, obtained from official government sources, were analyzed, along with apple area and production data. Exploratory interviews provided insights into installation costs, production losses, and perceived benefits. Results indicate a clear expansion of anti-hail nets associated with substantial hail losses. Producers reported pre-adoption losses of 30–90%, installation costs of Brazilian Real (BRL) 40,000–55,000 per hectare, and estimated payback periods of 3–4 years. These findings highlight the value of integrating remote sensing, official disaster records, and producer knowledge to support exploratory climate adaptation and economic assessments. 2026-01-01T00:00:00Z https://www.alice.cnptia.embrapa.br/alice/handle/doc/1188867 Título: Proprietary versus open-source visual-inertial fusion under GNSS degradation for orchard-scale 3D fruit mapping. Autoria: SANTOS, T. T.; BARTHI, D.; GEBLER, L.; RUFATO, A. de R. Conteúdo: A recent pipeline combining GNSS–visual-inertial odometry with factor-graph refinement of fruit landmarks has reported orchard-level apple counting errors below three percent against harvest totals — a result with few precedents in the agricultural SLAM literature, where GNSS-VIO fusion, landmark-level optimization, and harvest-validated yield estimation have until now appeared mostly in isolation. However, trajectory quality in that pipeline was assessed solely through reprojection error, which can decrease even when camera poses drift, provided poses and landmarks shift together consistently; low reprojection residuals therefore do not guarantee metrically accurate 3D fruit maps. Moreover, the pipeline relies on a proprietary sensor SDK whose internal fusion logic is opaque, making it impossible to attribute trajectory degradation to visual-inertial odometry quality, GNSS weighting, or downstream optimization. This study proposes a systematic evaluation protocol that jointly addresses both limitations. A postprocessed kinematic (PPK) GNSS signal is synthetically degraded before entering the fusion stage — through temporal subsampling, additive position noise, systematic drift, intermittent dropout, and a combined low-cost-receiver profile — while the original PPK trajectory is withheld as independent ground truth. Crucially, the degradation experiments are conducted on three parallel VIO-GNSS front-ends processing the same stereo imagery and IMU data from a tractormounted platform in a Fuji apple orchard in southern Brazil: the proprietary SDK fusion (ZED-F); an open OpenVINS estimator fused with degraded GNSS through an explicit GTSAM pose graph in which every factor covariance is transparent and controllable (OV-PG); and the same pose graph augmented with visual landmark projection factors (OV-PGL). All three fused trajectories feed the same downstream pipeline — YOLO-based apple detection, point-trackerbased temporal association, and multi-view triangulation — so that differences in trajectory error and counting accuracy are attributable to the fusion stage alone. For each front-end and degradation level, Absolute Trajectory Error and Relative Pose Error are computed against the withheld reference, and reprojection behavior is examined to identify the regime where internal consistency diverges from metric accuracy. The protocol is applied to two independent traversals of the same orchard block collected approximately one year apart, each with its own PPK ground truth. The results disentangle VIO front-end quality from fusion strategy: the open pose graph achieves the lowest absolute error where it completes, the proprietary fusion is the most robust across all conditions, and the landmark-augmented graph — although beneficial for local map rigidity — anchors the trajectory to corrupted GNSS once position noise exceeds a critical threshold, inflating absolute error by up to an order of magnitude while remaining internally consistent. Relative pose error proves insensitive to both architecture and degradation, confirming that internal or relative metrics alone cannot certify georeferenced fruit maps. Multi-view fruit triangulation recall is further shown to be a sensitive proxy for trajectory orientation accuracy, indicating that counting quality under realistic GNSS degradation is governed primarily by orientation consistency rather than by positional accuracy or GNSS signal quality. Together, these findings establish GNSS quality requirements for operational deployment of SLAM-based fruit mapping and contribute toward the reproducible multi-season evaluation protocols currently absent from the agricultural SLAM literature. 2026-01-01T00:00:00Z https://www.alice.cnptia.embrapa.br/alice/handle/doc/1188819 Título: Predicting carbon stocks in deeper soil layers using topsoil data. Autoria: MOURÃO, V. H. M.; KARATAY, Y. N.; BARIONI, L. G.; DAMIAN, J. M.; MOTTA, M. R. Conteúdo: ABSTRACT: Agricultural soils have a significant potential for carbon sequestration, thus playing a vital role in addressing climate change. Deeper soil layers, often overlooked in inventories, contain considerable amounts of soil organic carbon (SOC), particularly in tropical regions. Hence, it is crucial to include variations in those stocks in carbon accounting. However, the higher costs of measuring deep carbon stocks often deter such measurements. Therefore, developing cost-effective methods to predict SOC stocks in deeper soil layers is essential. This study aimed to assess the relationship between topsoil data and predictions of carbon stocks across various soil depths in tropical native vegetation and croplands on 53 farms in Brazil. We examined multiple combinations of soil layers above a target depth (e.g., 40 cm) to assess the viability of using topsoil data to predict deeper SOC stocks. Our results indicate that SOC stocks at depths of 30-40 cm and 40-60 cm can reliably predict SOC stocks at 40-100 cm and 60-100 cm, respectively. The models developed in this study provide a cost-effective approach for estimating SOC stocks in deeper soil layers, potentially enhancing the economic efficiency of quantifying the contributions of the agricultural sector and carbon farming initiatives in Brazil. 2026-01-01T00:00:00Z