Accurate biomass estimation remains critical for robotic harvesting, yet standard computer vision models severely underestimate crop size when broccoli heads are occluded by leaves. This paper investigates whether a high speed, edge viable modal model (YOLO26) can be trained to amodally segment occluded broccoli heads using targeted synthetic data instead of labor intensive manual annotations. To execute this, an automated processing pipeline superimposed leaves over fully visible crop backgrounds to synthesize realistic leaf occlusions. This generated synthetic dataset was subsequently deployed to train the network, establishing the data-driven amodal architecture STAR-YOLO26. The resulting model was evaluated against the heavier, two-stage ORCNN architecture across standard localization metrics, processing speed, and sizing parameters. The results indicate that STAR-YOLO26 successfully infers hidden crop boundaries producing comparable results to ORCNN with the primary differences being that STAR-YOLO26 systematically underestimates while ORCNN overestimates the shape area and that ORCNN performs better under heavier occlusions. Crucially, STAR-YOLO26 delivers these results while providing a massive 700% increase in processing speed compared to ORCNN. In conclusion, modal single-stage models demonstrate the capacity to generate amodal predictions when supplied with targeted training distributions, offering a precise, real-time, and scalable framework for automated crop segmentation and subsequent size estimation.


https://data.4tu.nl/datasets/aaef7e9c-47a5-44d4-af6c-1d2eb41b0f28
This dataset contains synthetically modified images based on the Data underlying the publication: Image-based size estimation of broccoli heads under varying degrees of occlusion. Version 3. 4TU.ResearchData. dataset. https://doi.org/10.4121/13603787.v3 by P.M. (Pieter) Blok; van Henten, Eldert; Frits van Evert; Gert Kootstra, used under CC BY-NC-SA 4.0. The original images were altered by overlapping leaves on fully visible broccoli heads to create varying occlusions (one leaf per image). The annotations included with these synthetically occluded images are the modal annotations of the original (fully visible) broccoli head, serving as the amodal mask for the model to train on. This dataset was used to train STAR-YOLO26, a YOLO26 model capable of amodal recovery of broccoli heads under occlusion, as part of a soon to be published bachelor thesis. In accordance with the ShareAlike clause of the original dataset, this derivative dataset is released under the same CC BY-NC-SA 4.0 terms. ... Read more

Broccoli farming requires accurate estimations of the head diameter to determine when crops are ready to be harvested. An existing growth prediction model was trained on data gathered at Verdonk Broccoli. This paper explores an alternative way of gathering the data needed to train the prediction model, namely by using an end-to-end deep learning model to estimate the diameter of a broccoli head based on an image and the camera height.

This new method was first evaluated by comparing the original diameter measurements with the model outputs. The impact of the method was then assessed by examining the accuracy of the prediction model when trained on the data generated using this approach.

The results show that this alternative method can achieve at least the same accuracy as the original approach, but further improvements in accuracy are limited by the dataset used to train the model. ... Read more

Broccoli head segmentation is the upstream step of a sizing pipeline that feeds harvest decisions, growth models, and yield forecasts, but the broccoli vision literature reports each on a different metric stack without confidence intervals, stratification, or robustness testing. This work identifies which evaluation variables actually differentiate broccoli head segmentation models under two deployment use cases: end-of-life-cycle harvest decisions and whole-cycle growth monitoring.

Five recent broccoli segmentation models are evaluated with cluster-bootstrap confidence intervals on two publicly available broccoli image datasets, under stratification by leaf occlusion, head size, and time of day; under two controlled photometric and motion-blur perturbations calibrated against published in-deployment image-sharpness anchors; and under a noise-injection sweep on a published downstream growth-prediction model.

Stratification by occlusion and by head size each flip the ranking between two of the broccoli segmentation models tested; no single architecture is operationally optimal across the growth curve. The two controlled perturbations agree on the qualitative robustness tiers but not on the strict rank order; a cross-camera evaluation produces no usable ordering at all. Upstream sizing-label noise propagates linearly into downstream growth-prediction MAE with a small coefficient, so cleaning sizing labels beyond the most accurate currently published sizing pipeline buys at most one to four hours of biological head growth, well under one harvest-decision cycle. The implication is that broccoli vision evaluation needs deployment-context-aware reporting rather than a single headline number, and sizing-pipeline accuracy is not the operational lever that moves downstream growth-prediction performance. ... Read more

Deciding when to harvest is an important part of broccoli horticulture; this is often done through sampling broccoli heads by hand. This process is both time-consuming and error-prone, making it a prime candidate for automation using computer vision models. This would make it feasible to estimate the size of all broccoli heads in a field.

Unfortunately, annotating the data required for a robust size estimation model is also a time-intensive and error-prone task. It is therefore interesting to look into ways in which image annotation can be streamlined.

This paper will show that using human-in-the-loop (HITL) annotation methods can speed up broccoli head annotation while minimally impacting the final segmentation model's segmentation and size estimation accuracy.

To achieve this, a dataset will be annotated three times, once manually, once with a graph cut-based segmentation algorithm known as GrabCut, and once with the prompted segmentation model SAM2. During this process the time taken will be recorded. After which the annotation models are trained on each set of annotations. This will be used to determine how each annotation method impacts a vision model's segmentation and size estimation accuracy after training.

The results show that HITL methods can speed up annotating, as using GrabCut shows a 3.1 times speed-up over manual annotation, and SAM2 shows a 6.2 times speed-up over manual annotation. The segmentation and size estimation capabilities of the methods post-training only differ by 1.5% when compared to each other, where manual annotation has the highest accuracy, followed by GrabCut.

This indicates a speed-accuracy trade-off for the chosen methods. Due to how small the accuracy degradation is, each method can be seen as worthwhile to us. Care should still be taken when deciding which method to use, as small errors can have greater effects downstream.
... Read more

Estimating the size of broccoli heads in the field is an important step towards automating broccoli harvesting, but it depends on first detecting the heads reliably and accurately in field imagery that contains harsh sunlight, camera blur, and leaf occlusion. This work investigates whether image preprocessing can improve such detection. Five common preprocessing techniques, CLAHE, unsharp masking, wavelet transform, median filtering, and bilateral filtering, were selected from the literature and evaluated individually and in five combinations. A systematic comparison was performed in which a YOLOv8n detector was trained and assessed with stratified five-fold cross-validation on a dataset of 394 field images. Performance was measured both over the entire dataset and over subsets representing difficult field conditions. On the full dataset the effect of preprocessing was negligible, as the baseline already performed close to the ceiling and differences between techniques were of the same order as the standard deviation between folds. However, in the difficult subsets preprocessing resulted in clear improvements, with the largest gains in the images on which the baseline struggled most. The most consistent results come from unsharp masking and the wavelet$+$unsharp combination. Combining techniques was not automatically beneficial. These results, obtained without significance testing, indicate that preprocessing is best understood as a tool for specific difficult conditions, rather than as a general improvement. ... Read more

Many of the most celebrated recent advances in artificial intelligence (AI) have been built on the back of highly complex and opaque models that need little human oversight to achieve strong predictive performance. But while their capacity to recognize patterns from raw data is impressive, their decision-making process is neither robust nor well understood. This has so far inhibited trust and widespread adoption of these technologies. This thesis contributes to research efforts aimed at tackling these challenges, through interdisciplinary insights and methodological contributions.

The principle goal of this work is to contribute methods that help us in making opaque AI models more trustworthy. Specifically, we aim to (1) explore and challenge existing technologies and paradigms in the field; (2) improve our ability to hold opaque models accountable through thorough scrutiny; and, (3) leverage the results of such scrutiny during training to improve the trustworthiness of models. Methodologically, the thesis focuses on counterfactual explanations and algorithmic recourse for individuals subjected to opaque AI systems. We explore what type of real-world dynamics can be expected to play out when recourse is provided and implemented in practice. Based on our finding that individual cost minimization–a core objective in recourse–neglects hidden external costs of recourse itself, we revisit yet another established objective: namely, that explanations should be plausible first and foremost. Our work demonstrates that a narrow focus on this objective can mislead us into trusting fundamentally untrustworthy systems. To avoid this scenario, we propose a novel method that aids us in disclosing explanations that are maximally faithful, that is consistent with the behavior of models. This not only allows us to assess the trustworthiness of models, but also improve it: we show that faithful explanations can be used during training to ensure that models learn plausible explanations.

Finally, we also critically assess efforts towards trustworthy AI in the context of modern large language models (LLM). Specifically, we cast doubt on recent findings and practices presented in the field of mechanistic interpretability and caution our fellow researchers in this space against misinterpreting and inflating their findings.

In summary, this thesis makes cutting-edge research contributions that improve our ability to make opaque AI models more trustworthy. Beyond our core research contributions, this thesis makes substantial contributions to open-source software. Through various software packages that we have developed, we make our research and that of others more accessible.
... Read more

Water flows through every aspect of life, yet the story of its delivery is only as reliable as the data that records it. In global benchmarking, such data is often uneven, incomplete, and rarely subjected to systematic validation, allowing anomalies to shape perceptions of performance before they are critically examined. This thesis addresses that gap by developing and evaluating a multi‐stage, data‐driven anomaly detection framework within the World Bank’s New International Benchmarking Network for Water and Sanitation Utilities (NewIBNET), situated at the intersection of data science, water governance, and digital ethics.

The framework weaves together four complementary layers – structural validation, rule‐based logical checks, peer comparison, and weighted prioritisation – transforming anomaly detection from a surface‐level cleaning task into a structured process of active quality assurance. Developed through an iterative, expert‐informed process, it is reproducible and adaptable, balancing statistical rigour with the contextual realities of the water sector so that each flag raised carries both analytical credibility and practical relevance.

Applied to the 2022–2024 NewIBNET dataset, the framework is assessed through robustness checks, a national case study of Indonesian utilities, and an expert survey. Results show that it improves anomaly interpretability, limits the propagation of flawed data into comparative analyses, and reduces review time from 75 hours to under 2 minutes – earning unanimous expert endorsement for operational deployment.

By translating the principles of automated, ethically grounded validation into a scalable methodology, this work advances the state of practice in anomaly detection for data‐scarce sectors. In shifting from red flags to real solutions, it demonstrates how automated validation can turn detection into action, building trust where data meets water, and enabling more transparent, equitable decisions in global water governance. ... Read more

This study examines dataset annotation practices in influential NeurIPS research. Datasets employed in highly cited NeurIPS papers were assessed based on criteria concerning their item population, labelling schema, and annotation process. While high-level information, such as the presence of human labellers and item population, is present in most cases, procedural details of the annotation process are poorly reported. Notably, 48% of datasets lack details on annotator training, 43% omit inter-rater reliability, and 28% are not publicly accessible. Temporal comparisons show minor improvements, but no substantial progress in reporting annotation methodology. A complementary analysis of 49 NeurIPS papers published since 2020 shows that researchers often discuss the broader impact of their work, yet do not include datasets or their annotations in these assessments. These findings highlight a lack of standardisation in annotation reporting and call for more robust practices that ensure transparency, auditability, and reproducibility in machine learning research. ... Read more

As humanitarian needs increase while donor budgets decrease, anticipatory strategies are essential for effective crisis response. In this context, machine learning (ML) has emerged as a promising tool for crisis forecasting, offering the potential to support timely interventions and humanitarian decision-making. However, despite rapid developments in ML-based prediction models, questions remain about their practical utility and trustworthiness in real-world humanitarian settings. This study presents a systematic scoping review of 32 academic and gray literature sources to assess the reliability and feasibility of ML systems for conflict forecasting. By analyzing these systems across dimensions such as forecasting scope, data sources, modeling approaches, validation practices, and ethical considerations, the study finds that while some models demonstrate strong predictive performance and methodological rigor, many lack transparent validation, robust error analysis, and operational applicability. The review concludes that while ML systems hold substantial potential for enhancing conflict anticipation, their current real-world readiness is uneven and context-dependent. ... Read more

Displacement is a focal point of humanitarian aid efforts, since it affects millions of people globally. Mitigating the consequences of forced migration is important for reducing suffering and one way of doing so is through predicting displacement to prioritise resources in advance. To achieve this, machine learning can be used for its ability to analyse larger amounts of data and identify latent structures more efficiently than human experts. Through a systematized literature review, this research evaluates thoroughly six machine learning tools: UNHCR's Jetson, DRC's Foresight and AHEAD, the EU's EUMigraTool and EPS-Forecasting, and the agent-based simulation Flee, in order to assess their suitability to that end. The analysis compares these tools across several criteria, including the way they use data, algorithmic characteristics, and operational use cases. Finally, it makes recommendations about what should be considered and how to choose amongst the tools for displacement prediction. ... Read more

Natural disasters frequently cause casualties and property losses. Predicting and mitigating the impact of such threats is crucial to the work of humanitarian organizations. The interactions between hazards are best represented through a multi-hazard approach, and machine learning models are well suited for natural hazard prediction. This study presents a systematized literature survey of machine learning in multi-hazard disaster forecasting in the years 2019-2025, focusing on the used models and performance metrics, their applications and feasibility of use, as well as potential cross-applications. There is a wide variety of models and metrics used. The most commonly used models are random forest and support vector machine and the most prevalent performance metric is the ROC-AUC score. The machine learning models generally perform well, with AUC scores above 0.8, though patterns in performance are difficult to examine. Feasibility is defined here as readiness to be used in practice, and the models are rated in the factors that define it. Most of the articles are feasible. Consideration of cross-application is rare and should be extended. This research summarizes the main trends in the field of disaster forecasting, providing a clear reference point for other academics. ... Read more

The output of machine learning (ML) models can be only as good as the data that is fed into them. Because of this, when making datasets for creating ML models, it is important to ensure the quality of the data. This is especially true of human labeled data, which can be hard to standardize and assess the quality of. To assess the annotation practices of human labeled data in the field of machine learning, this paper investigates the datasets used in the highest cited papers in the AAAI Conference on Artificial Intelligence, an influential machine learning conference. After extracting the datasets from 75 papers in three overlapping publication periods, the top 20 datasets were evaluated from each period. The results showed that the majority of datasets do not use or underreport significant annotation practices, specifically about the annotators and the annotation process. This raises concern for the conference and the field more broadly, as the most influential papers build their machine learning algorithms on quite possibly low quality data. However, there is some hope for the field in this regard as the more recent papers use datasets with better quality annotation practices. ... Read more

With the worsening of climate change, the complications brought on by floods every year create an increasing need for forecasting systems that humanitarian organizations can use to help populations in danger. This research presents a literature review of machine-learning models for impact-based flood forecasting, and compares them with existing humanitarian projects. The results examine the characteristics of the models surveyed, while the discussion focuses on understanding how these characteristics can define whether the machine learning models proposed can actually be translated to humanitarian settings. The main takeaways include the prevalent choice of deep learning and ensemble models, used to improve the adaptability of the models, the problems with data availability and data quality in different areas considered, and the difference between lead times, usability, and scalability of the models proposed in contrast with already used humanitarian projects. This study then highlights the importance of transparency and reproducibility of the survey by detailing the queries and databases used, ensuring accessibility of selected articles, and explaining the selection criteria and methodology. Ultimately, the review concludes with the key insights on the connection between academic prototypes and real-life humanitarian projects, as well as key areas for future research. ... Read more

This study gives an overview of the data collection and annotation practices of the datasets used by the most impactful papers published by the Association of Computational Linguistics (ACL). This was achieved by selecting the most highly cited papers published within the ACL anthology across 3 periods (published in the past 2, 5 and 15 years). Afterwards, the datasets used by those papers were extracted and filtered to retain the most impactful ones. Finally, a carefully crafted annotation schema was used to find out information regarding key aspects of the datasets in order to qualitatively analyze them. As a result of this analysis, it was first found that (1) there are fewer datasets used on average in the past 2 years and that there is little overlap with the datasets used by papers published in the past 5 or 15 years. (2) Secondly, there are various concerns related to those key aspects, such as the relatively high (∼36%) and unregulated use of the Amazon Mechanical Turk crowdsourcing platform for the construction of datasets. Another concern is information frequently missing about any rationale regarding labeller population, prescreening, inter-rater reliability and rationale regarding sample size - missing ∼77%, ∼63%, ∼19-56%, and ∼81% of the time. However, reporting practices for most of those issues have slightly improved within datasets used in the past 2 years. (3) Finally, around one third of the information sought was missing across all periods. However, the state of the domain has been generally improving, with a lower one fourth of the information missing from datasets used in the past 2 years. Some recommendations are given in order to overcome those challenges, the most important of which being that each academic organization should require their submissions to include a reporting template in their papers. ... Read more

High-impact vision research still rests on datasets whose labels arrive via opaque, rarely documented pipelines. To understand how serious the problem is inside a large venue, we audited 75 TPAMI papers (2009-2024) that rely or introduce datasets. Each dataset was coded against a 27-item checklist adapted from Garbage in, Garbage out, spanning annotator recruitment, training, compensation, overlap-resolution and more. Across the corpus, 37% of the expected annotation metadata is missing; the rate changes little between recent (2022-24) and older cohorts. The scarcest fields are labeller-population rationale (76.6% absent), prescreening criteria (73.4%), total annotators (68.8%), compensation (67.2%) and training procedures (62.5%). Documentation quality shows virtually no correlation with a paper’s citation impact, suggesting community prestige does not buy transparency. A handful of well—curated datasets achieve >75% completeness, proving that thorough documentation is possible when incentives align. The median TPAMI benchmark still ships with an unverifiable "ground truth", threatening the reproducibility and fairness claims of downstream models. We advocate that journals and conferences require a concise, checklist-based annotation statement, mirroring existing ethics and reproducibility forms, to ensure future vision systems are built (and evaluated) on transparent, trustworthy data foundations. ... Read more

This study investigates annotation and reporting practices in machine learning (ML) research, focusing on societally impactful applications presented at the IEEE/CVF Computer Vision and Pattern Recognition (CVPR) conferences. By structurally analyzing the 75 most-cited CVPR papers from the past 2, 5, and 15 years, we evaluate how the human annotations foundation of supervised ML is documented. We introduce a 27-field annotation-reporting schema and apply it to 60 datasets, revealing that nearly 30% of relevant information is routinely omitted. Key findings include the pervasive underreporting of annotator details such as training, prescreening, and inter-rater reliability (IRR) metrics. While popular datasets like COCO and ImageNet exhibit widespread use, transparency about annotation methodologies remains inconsistent. The impact of a few fields shows that basic metadata, such as the selection process of annotators and how the labels' overlap is managed, strongly anticipate overall documentation quality. Our findings support previous calls for standardization and underscore the need for institutionalized reporting practices to ensure reproducibility, fairness, and trust in ML systems. ... Read more

In musical (jazz) improvisation, musicians that are just starting out can often feel uncomfortable when being put on the spot by their fellow players. However, when a musician is on their own when practising or leisurely playing, this prevents them from listening to fellow musicians. When a musician wants to experience some notion of co-play when they are on their own, computers and musical generative techniques may be a source of help. We study the extent to which music improvisation algorithms can facilitate such interactions by proposing an experimental framework to evaluate and compare these different algorithms. We achieve this by developing MILES ('Mixed-Initiative musicaL interactivE System'), a generic music improvisation system that allows a musician to improvise with various musical improvisation models and facilitates comparative evaluation of these models. MILES makes use of the 'trading fours' paradigm, where two or more musicians exchange four measures of solo material. We conduct experiments with novice and advanced musicians in expert-pupil and peer-to-peer settings that compare differing algorithms, as well as different variations of similar algorithms. These comparisons are based on self-assessed opinions and third-party grading and ordering of recordings, based on improvisational reciprocity and enjoyment. Symbolic music recording analysis further quantifies the interactivity between the musician and the algorithms. With this experimental setup, we are able to track familiarity and enjoyment of using music improvisation algorithms, and compare different iterations of similar music improvisation algorithms. ... Read more

Machine Learning (ML) systems are increasingly used in high-stakes, socially impactful domains, requiring attention to improve explainability and trust. However, current Requirements Engineering (RE) techniques often fail to address these human-centered qualities. This research investigates how Behavior-Driven Development (BDD) and Goal-Oriented Requirements Engineering (GORE) can improve the identification and visualization of requirements for explainability and trust. We conducted expert interviews and a survey to see how different stakeholders rate certain BDD scenarios and what they think of conceptual GORE models. Our results show that participants value concise, human-readable BDD scenarios and particularly like the GORE framework of i* to understand stakeholder relationships and system behaviors. The other GORE framework, GR4ML, was found to align more with business goals and addresses other stakeholder perspectives less. We conclude that BDD and GORE can improve explainability in ML system development. Future work should refine modeling tools to better integrate ethical and fairness considerations. ... Read more

Broccoli for the fresh market in the Netherlands is still harvested manually, which is labor-intensive and increasingly difficult to sustain as seasonal labor declines. Existing mechanized harvesters cut entire fields at once and cannot account for plant-to-plant variation, leading to substantial losses when heads differ in maturity. These constraints motivate plant-level, data-driven growth modeling to support selective and more efficient harvesting.

This thesis investigates how field-based measurements can be used to model broccoli growth at the level of individual plants. The work addresses four problems: converting raw field video into plant-level growth curves, investigating which environmental parameters best describe the growth, determining whether cumulative temperature or thermal time better represents broccoli development, and evaluating how different growth models capture head diameter growth. A preliminary study using an external dataset creates the methodological foundation by analysing broccoli growth and benchmarking classical and neural models. A field study in a Dutch production environment expands this work through the development of a data-processing pipeline that includes head detection, plant identification, tracking, diameter estimation, and integration with local weather data. Across both studies, thermal time provides a more biologically meaningful predictor of development than cumulative temperature. Model comparison shows that classical parametric models capture general developmental trends, while a multi-layer perceptron achieves the highest predictive accuracy, with a mean absolute error of 0.583~cm, when multiple environmental variables are included. The results demonstrate that field-based predictive modeling can support precision-agriculture applications such as harvest planning and selective mechanized harvesting. ... Read more

Central banks communicate their monetary policy plans to the public through meeting minutes or transcripts. These communications can have immense effects on markets and are often the subjects of studies in the financial literature. The recent advancements in Natural Language Processing have prompted researchers to analyze these communications using Transformer-based Large Language Model (LLM) classifiers. The use of LLMs in finance and other high-stakes domains calls for a high level of trustworthiness and explainability of those models. We focus on Counterfactual Explanations, a form of Explainable AI that explains a model's classification by proposing an alternative to the original input. We use three types of CE generators for LLM classifiers on a recent dataset consisting of sentences taken from FOMC communications to assess the usability of their explanations. We perform three experiments comparing different types of generators, one using a selection of quantitative metrics and two involving human evaluators, including central bank employees. Our findings suggest that non-expert and expert evaluators prefer counterfactual methods that apply minimal changes to the texts; however, the methods we analyze might not handle the domain-specific vocabulary well enough to generate plausible explanations for our task. We discuss shortcomings in the choice of evaluation metrics in the literature on text CE generators and propose refined definitions of the fluency and plausibility qualitative metrics. ... Read more