Organizations: The National School of Artificial Intelligence (ENSIA), Algiers, 16046, Algeria · National Higher School of Agronomy (ENSA), Algiers, 16051, Algeria · National Higher School of Computer Science (ESI), Algiers, 16000, Algeria · National Higher School of Advanced Technologies (ENSTA), Algiers, 16000, Algeria
Machine learning models in agricultural vision often achieve high accuracy on curated datasets but fail to generalize under real field conditions due to distribution shifts between training and deployment environments. Moreover, most machine learning competitions focus primarily on model design while treating datasets as fixed resources, leaving the role of data collection practices in model generalization largely unexplored. We introduce the AgrI Challenge, a data-centric competition framework in which multiple teams independently collect field datasets, producing a heterogeneous multi-source benchmark that reflects realistic variability in acquisition conditions. To systematically evaluate cross-domain generalization across independently collected datasets, we propose Cross-Team Validation (CTV), an evaluation paradigm that treats each team's dataset as a distinct domain. CTV includes two complementary protocols: Train-on-One-Team-Only (TOTO), which measures single-source generalization, and Leave-One-Team-Out (LOTO), which evaluates collaborative multi-source training. Experiments reveal substantial generalization gaps under single-source training: models achieve near-perfect validation accuracy yet exhibit validation-test gaps of up to 16.20% (DenseNet121) and 11.37% (Swin Transformer) when evaluated on datasets collected by other teams. In contrast, collaborative multi-source training dramatically improves robustness, reducing the gap to 2.82% and 1.78%, respectively. The challenge also produced a publicly available dataset of 50,673 field images of six tree species collected by twelve independent teams, providing a diverse benchmark for studying domain shift and data-centric learning in agricultural vision.
Compound, pinnate; 7–11 leaflets; foliage turns red in autumn
Smooth grey-green trunk; dense, compact crown
Table 3: Visual and morphological properties of the six target species.
Figure 2: Samples from the AgrI Challenge dataset across the six target tree species.
Team
Carob tree
Oak
Peruvian pepper tree
Ash
Pistachio tree
Tipu tree
Total
AI-4o
960
1625
1221
934
1111
1493
7344
AiGro
686
604
595
572
541
635
3633
CACTUS
400
597
694
803
755
552
3801
CHAJARA
610
549
428
308
262
383
2540
GreenAI
675
705
609
559
722
515
3785
PLT
1089
1131
1000
1297
973
1066
6556
Table 4: Distribution of images across teams and tree species classes
Figure 3: Histogram showing the distribution of dataset images by capture device model. The dataset contains images from more than 40 device types, with the largest groups originating from iPhone 11 and Oppo Reno5, and a substantial portion labeled as Unknown due to missing metadata.
Model
Type
Parameters
Purpose
DenseNet121
CNN
8M
Efficient backbone suitable for resource-constrained scenarios and strong feature reuse.
Swin-Tiny
Transformer
28M
Lightweight transformer baseline capturing both local and global contextual information through window-based attention.
Table 5: Baseline model architectures used for comparison.
Parameter
Value
Hardware
NVIDIA H100 NVL (22 GB)
Pre-training
ImageNet-1K
Optimizer
AdamW
Weight decay
10−4
Initial learning rate
10−4
Learning rate schedule
Cosine annealing ( ηmin=10−6 )
Table 6: Training configuration used for all experiments.
Team
DenseNet121
Swin Transformer
Val Acc.
Test Acc.
VTG
Val Acc.
Test Acc.
VTG
AI-4o
96.60
86.94
9.65
98.09
91.08
7.02
AiGro
98.99
82.32
16.67
99.27
87.68
11.58
CACTUS
96.84
84.36
12.49
98.86
90.19
8.67
CHAJARA
97.51
81.56
15.94
98.82
88.03
10.79
GreenAI
96.92
85.47
11.45
98.42
90.86
7.55
Table 7: Controlled TOTO evaluation with standardized architectures
Figure 4: Cross-team test accuracy for DenseNet121 and Swin Transformer in TOTO protocol. Horizontal dashed lines indicate means: 81.19% (DenseNet) and 87.21% (Swin).
Figure 5: Global model performance under the TOTO protocol for DenseNet121 and Swin Transformer (mean ± std across 12 runs). The large validation–test gap highlights the cross-team generalization challenge under single-team training.
Figure 6: Aggregate learning curves for TOTO protocol ( n=12 runs, mean ± std). Left: Swin Transformer. Right: DenseNet121. The gap between validation and test accuracy persists across all epochs.
Figure 7: Distribution of validation–test gaps under the TOTO protocol for Swin Transformer and DenseNet121. The red dashed line marks zero gap. Larger gaps indicate stronger cross-team generalization loss.
Figure 8: DenseNet121 cross-team accuracy matrix. Rows: training teams; columns: test teams. Range: 48.2%–95.3%.
Table 8: LOTO evaluation using standardized DenseNet121 and Swin Transformer architectures.
Figure 10: Distribution of validation-test gaps under LOTO protocol for Swin Transformer and DenseNet121. The red dashed line marks zero gap. Two teams achieved negative gaps in both models, indicating the held-out data was easier to generalize to than the validation set. The single outlier (DenseNet, Condimenteum: +10.76%) reflects that team’s distinctive collection methodology.
Figure 11: Global model performance under LOTO protocol for DenseNet121 and Swin Transformer (mean ± std across 12 folds). The compressed train-validation-test spread contrasts sharply with the large gaps observed under TOTO.
Figure 12: Aggregate LOTO learning curves. Left: Swin Transformer (mean final test: 97.04%). Right: DenseNet121 (mean final test: 95.31%).
Figure 13: Per-held-out-team test accuracy under LOTO protocol for Swin Transformer (left) and DenseNet121 (right). Dashed blue lines indicate overall means (97.04% and 95.31%). The consistency of team rankings across both architectures (Spearman ρ=0.92 ) confirms that residual variation reflects dataset-level properties rather than architecture-specific effects.
Modern agricultural data is sourced from diverse platforms and spans multiple spatial scales, ranging from ground-level close-up photography to Unmanned Aerial Vehicle (UAV) aerial observation and satellite remote sensing imagery. Accordingly, agricultural multimodal reasoning demands robust cross-scale spatial understanding. However, due to the lack of multi-view agricultural benchmark datasets, existing multimodal large language models (MLLMs) exhibit severe ground-level bias, which leads to scale confusion then semantic collapse in agricultural perception tasks, such as misinterpreting farmland imagery as walls or floors. To address this, we introduce AgroOmni, a large-scale multi-view training corpus with 288K Visual Question Answering pairs covering 56 specialized task categories across 14 task types, designed to capture diverse scales in modern precision agriculture. Built on this dataset, we propose AgroNVILA, which achieves a new state-of-the-art of 62.32% on the AgroMind benchmark (+15.03% over GPT-5.2), effectively mitigating the multi-view cross-scale gap for holistic agricultural understanding. Diagnostic evaluations on AgMMU further reveal an inherent heterogeneity between macro-priors and micro-diagnostics through constrained zero-shot performance. Meanwhile, even minimal fine-tuning leads to a dramatic performance gain of AgroNVILA on AgMMU, strongly demonstrating its generalization capability empowered by AgroOmni. Full training scripts are publicly available at https://anonymous.4open.science/r/AgroOmni-6510.
Jiarui Zhang, Junqi Hu, Zurong Mai +10
Sun Yat-sen University · The Chinese University of Hong Kong · HuanTian Wisdom Technology Co., Ltd. +4
The deployment of Multimodal Large Language Models (MLLMs) in agriculture is currently stalled by a critical trade-off: the existing literature lacks the large-scale agricultural datasets required for robust model development and evaluation, while current state-of-the-art models lack the verified domain expertise necessary to reason across diverse taxonomies. To address these challenges, we propose the Vision-to-Verified-Knowledge (V2VK) pipeline, a novel generative AI-driven annotation framework that integrates visual captioning with web-augmented scientific retrieval to autonomously generate the AgriMM benchmark, effectively eliminating biological hallucinations by grounding training data in verified phytopathological literature. The AgriMM benchmark contains over 3,000 agricultural classes and more than 607k VQAs spanning multiple tasks, including fine-grained plant species identification, plant disease symptom recognition, crop counting, and ripeness assessment. Leveraging this verifiable data, we present AgriChat, a specialized MLLM that presents broad knowledge across thousands of agricultural classes and provides detailed agricultural assessments with extensive explanations. Extensive evaluation across diverse tasks, datasets, and evaluation conditions reveals both the capabilities and limitations of current agricultural MLLMs, while demonstrating AgriChat's superior performance over other open-source models, including internal and external benchmarks. The results validate that preserving visual detail combined with web-verified knowledge constitutes a reliable pathway toward robust and trustworthy agricultural AI. The code and dataset are publicly available at https://github.com/boudiafA/AgriChat .
Abderrahmene Boudiaf, Irfan Hussain, Sajid Javed
Department of Computer Science, Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates
Field-based agricultural computer vision is important for precision agriculture, yet it largely depends on expensive annotations and costly adaptation of large pretrained models. We introduce AgriField-40K, a field-centric dataset curated from 17 public resources and covering diverse crops, weeds, and field conditions. Building on this, we present AgriMAE, a parameter-efficient continual pretraining baseline that adapts a masked autoencoder pretrained on natural images by training only lightweight adapters. We further explore semantic feature reconstruction as an alternative pretraining objective and evaluate transfer across multiple tasks. AgriMAE consistently improves downstream performance and can match or even outperform full fine-tuning while using up to 9× fewer trainable parameters, showing that AgriField-40K is a practical resource for continual pretraining in agricultural vision. Project page: https://dtu-pas.github.io/agrifield40k/