Publications

We propose ST-LoRA, a parameter-efficient ensembling method that combines Low-Rank Adaptation with snapshot ensembling for agricultural semantic segmentation, keeping the pretrained backbone frozen and training under 10% of a full model's weights. Evaluated on GrowliFlower-L (cauliflower) and BUP20 (sweet pepper) with SegFormer and Mask2Former, ST-LoRA matches or exceeds full-rank ensembles while cutting compute, training time, and memory, and shows consistent gains in out-of-distribution detection and calibration over baselines such as Snapshot Ensemble, MC Dropout, and Deep Deterministic Uncertainty.

We introduce an annotated image time-series dataset of 691 sweet pepper plants over two growing seasons (4,837 images with per-plant fruit counts by maturity) and a multimodal deep learning framework that fuses DinoV3 image features with numerical counts through an LSTM to model temporal dependencies under irregular sampling. The approach reduces RMSE over a persistence baseline by 33–38% across seasons, and we use Deep Ensembles with Gaussian NLL to provide calibrated uncertainty estimates for real-world yield forecasting.

We present two robot-collected datasets for visual monitoring of tomato plants in commercial-style greenhouses: BUTom21, static images with manual pixel-level annotations, and BUTom-ST21, video-based data with semi-automated pseudo-labels. Both provide pixel-level fruit-ripeness labels, and the spatial-temporal set adds per-fruit counts and ripeness tracking over time, giving researchers realistic imagery for detection, segmentation, tracking, and video-instance segmentation in field-based phenotyping.

We tackle automatic detection of unknown materials (e.g. fertilizer, sand, soil) and surface anomalies (e.g. cracks, holes) in multispectral imagery without labeled anomaly examples during training. The self-supervised and unsupervised methods we study are motivated by autonomous systems and environmental monitoring that need to flag unknown or hazardous substances — such as chemical spills, fuel leaks, or agricultural runoff — to prevent accidents and ecological damage.

This paper investigates conformal prediction as a framework for uncertainty quantification in crop production systems. We analyze how conformal prediction can enhance decision support by providing reliable prediction sets with coverage guarantees, comparing it against established uncertainty methods across various agricultural monitoring tasks.

We propose the Confident Naturalness Explanation (CNE) framework that combines explainability methods with uncertainty-aware confidence measures to explain and assess the patterns that form naturalness in remote sensing imagery. The framework bridges the gap between model interpretability and ecological assessment.

NaT-ReX is a transformer-based explainability framework for assessing naturalness in satellite imagery, combining explainable ML with uncertainty-aware confidence to highlight regions reliably associated with naturalness while discounting high-uncertainty areas. It introduces the ReX score, a pixel-wise metric relating land-cover classes to their contribution to naturalness, and finds that shrublands and wetlands score highest while open water and snow-dominated regions score lowest — consistent with expert-informed assessments.

This work presents a comprehensive framework for integrating explainable machine learning into digital agriculture decision support systems. We demonstrate how explainability methods can enhance trust and understanding in crop monitoring predictions, evaluated across multiple agricultural datasets.

We present a comparative study of uncertainty quantification methods for harvest-readiness classification in cauliflower plants. Our evaluation focuses on inductive conformal prediction, comparing its calibration properties, computational efficiency, and coverage guarantees against ensemble-based and Bayesian uncertainty methods.

This paper proposes an automatic severity classification system for diabetic retinopathy using DenseNet enhanced with Convolutional Block Attention Module (CBAM). The attention mechanism improves the network's ability to focus on relevant retinal features, achieving competitive classification accuracy across multiple severity levels.