Conference Agenda

Cucumber is an essential economic crop widely cultivated in summer. The hot and humid conditions make cucumber highly susceptible to diseases, pests, and disorders (DPD). Early and accurate detection is crucial to limit the spread of DPD. Conventional DPD identification relies on visual inspection or laboratory analysis, which are often subjective, labor-intensive, and time-consuming. Moreover, DPD often co-occur on a single leaf, further increasing the difficulty of diagnosis in real-world. To address these challenges, this study proposes to identify cucumber leaf DPD using multimodal and multi-label learning approaches. The dataset contains 13,473 cucumber leaf images and covers 10 DPD categories. The label cardinality of the dataset was 1.229, reflecting the frequency of label co-occurrence. The framework adopts a two-stage design. First, a teacher–student anomaly detection module with a CvT-13 backbone and SVM classifier filters out non-cucumber images. Second, a multimodal and multi-label identification framework was developed. A vision transformer extracts visual features, while BERT encodes textual symptom descriptions provided with the image. The visual and textual semantic features were fused to capture co-occurring symptom patterns and independently predict multiple DPD categories in an end-to-end manner. The proposed framework supports practical cucumber leaf DPD identification in real-world settings.

Detecting microcracks on highly textured agricultural products such as Yunnan deep-texture walnuts remains a significant challenge due to extreme visual camouflage and weak structural contrast, directly affecting post-harvest quality and safety. This study proposes a semi-supervised inspection framework (Semi-WSCNet) for accurate and cost-efficient microcrack detection under industrial conditions. An Adaptive Frequency Decoupling module separates texture-dominated backgrounds from crack-sensitive components through learnable wavelet bases, while an Attention-Guided Feature Aggregation module enhances multi-scale crack continuity perception. A morphology-aware refinement mechanism further improves structural consistency. To reduce annotation burden, an uncertainty-guided semi-supervised strategy progressively optimizes detection performance using pseudo-labels.

Experiments on a challenging deep-texture walnut dataset demonstrate that the proposed system detects microcracks narrower than 50 μm, achieving a Structure Measure of 0.9303 and a Mean Absolute Error of 0.0012. Compared with the fully supervised baseline, annotation cost is reduced by 78% while recall improves by 14.91%. The system operates at 37 frames per second, satisfying real-time industrial requirements.

The proposed framework provides a practical and scalable solution for automated walnut sorting and advances semi-supervised inspection for camouflaged defects in complex agricultural environments.

This study aimed to identify the most relevant input variables for estimating two peanut maturity indices using remote sensing data and artificial neural networks, with an emphasis on feature selection to reduce multicollinearity and input dimensionality. The experiment was conducted in a commercial peanut field in São Paulo State, Brazil, using the cultivar IAC 503. PlanetScope satellite imagery and field measurements were collected on the same day. Candidate predictors included growing degree days, spectral bands, vegetation indices, and topographic indices. A variable selection procedure was applied to identify the most informative predictors and remove redundant information before model development. The selected variables were then used to train and test two types of artificial neural network models: Multilayer Perceptron and Radial Basis Function networks. Results showed that a reduced set of input variables was sufficient to estimate both peanut maturity indices, achieving R² values of 0.91 and 0.96 and MAE values of 0.05 and 0.06, respectively. The proposed approach demonstrated robust and efficient predictive performance, highlighting the importance of feature selection prior to machine learning modeling. In addition, it provides decision support for defining the optimal harvest time, contributing to loss reduction and improved crop management in precision agriculture.

Monitoring insects in natural environments requires species identification methods. Microphone-based systems offer a promising approach, as flying insects produce characteristic wingbeat sounds that differ between species and can be used for automated classification.

Training DL models for insect sound classification, however, typically requires large labeled datasets and computational resources. Meanwhile, several pretrained acoustic models exist, trained on general audio or bird vocalizations. It remains unclear whether these models provide useful feature representations for insect sounds.

We evaluated pretrained acoustic models as feature extractors for insect classification using the public InsectSound1000 dataset, containing 169,000 recordings from 12 species. Feature embeddings extracted from multiple pretrained models were analyzed using UMAP and evaluated through downstream classification with a multi-layer perceptron.

Results show that models pretrained on general audio perform comparably to models trained on biological domains. For example, embeddings from the Perch_bird model achieved an accuracy of 52.9%, similar to BYOL-A (53.6%), with class accuracies exceeding 70% for several species.

These findings indicate that pretrained acoustic models, even from distinct domains, can capture meaningful representations of insect sounds, enabling rapid exploration of feature extractors while reducing the need to train large deep networks from scratch and lowering dependence on large labeled datasets.

Olive harvest timing critically influences both oil yield and quality. Effective harvest planning involves assessing the Olive Maturity Index (MI) and the load of hanging fruit. The traditional MI method developed at the Experimental Station of Venta del Llano is based on the visual classification of epicarp and mesocarp pigmentation into eight ripening categories. Although this method is widely used, it can be challenging to apply consistently across large, diverse orchards.

In contrast, the five-class color-based scale proposed by Guzman et al. aligns better with modern early-harvest strategies and is especially well-suited for automated image analysis.

This study evaluates a non-destructive, in vivo, continuous RGB imaging system for estimating maturity and yield in a large, high-density olive orchard. A vehicle equipped with an automated image acquisition system was used to gather georeferenced RGB images along crop rows, while a. fully automated computer vision pipeline computed MI and estimated the fruit load.

The RGB-derived color index showed a strong correlation with the MI, validating the automated approach. The automatically computed MI closely matched the visual one, while the image-based fruit counts showed a strong predictive correlation with actual yield. This system offers a scalable and objective solution for precision harvest management.