Durian: Crop Insights & RL Normalization

• Durian is a dual-meaning term, denoting both an agricultural crop studied for disease recognition and a normalization method in multimodal reinforcement learning.
• Precision agriculture studies on durian address real-world challenges using Mixture-of-Experts and IoT-based techniques for disease detection and animal intrusion prevention.
• The Durian RL method enhances multimodal reasoning by regrouping samples based on perceptual and reasoning difficulty, improving stability and performance.

Durian appears in recent machine-learning literature in two distinct but technically connected senses: as a crop domain for precision-agriculture systems, and as the acronym of “difficulty-aware group normalization (Durian)” in multimodal reinforcement learning. As a crop, durian is used to study plant disease classification under real-field conditions and orchard protection against wildlife intrusion, with datasets and deployment scenarios drawn from Vietnam and Malaysia (Hoang et al., 22 Jun 2026, Shahrir et al., 2 Nov 2025). As an algorithmic acronym, Durian denotes a modification to RLVR and GRPO for multimodal LLMs, designed to stabilize std-based normalization by regrouping samples according to perceptual complexity and reasoning uncertainty (Li et al., 25 Feb 2026).

1. Durian as a precision-agriculture research domain

In the agricultural literature considered here, durian is a case study crop for two separate but complementary operational problems: disease recognition from field imagery and plantation protection against animals. In "Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification" (Hoang et al., 22 Jun 2026), the durian disease dataset is one of three evaluation datasets and serves to assess cross-dataset generalization of a Mixture-of-Experts framework. In "A Hybrid YOLOv5-SSD IoT-Based Animal Detection System for Durian Plantation Protection" (Shahrir et al., 2 Nov 2025), durian plantations are the deployment environment for an IoT-based edge-vision system targeting elephant, boar, and monkey intrusion.

The two problem settings share a common framing. Plant disease classification is presented as critical for crop protection, yield, precision agriculture, and large-scale monitoring, with the explicit goal of replacing or supporting human visual inspection in the field (Hoang et al., 22 Jun 2026). Plantation protection is motivated by the inability of manual patrolling, noise makers, and simple sensors to provide continuous monitoring, animal type recognition, or rapid remote response (Shahrir et al., 2 Nov 2025). The resulting research agenda places durian within a broader smart-farming pipeline in which image acquisition, model inference, and operational intervention are all embedded in real orchard conditions rather than laboratory settings.

A common misconception is to treat durian in these studies as merely a botanical label. The cited work instead uses durian as an operational domain with specific sensing, inference, and deployment constraints: natural lighting, cluttered backgrounds, smartphone capture, edge inference, real-time notification, and, in the wildlife case, automated deterrence (Hoang et al., 22 Jun 2026, Shahrir et al., 2 Nov 2025).

2. Durian disease dataset and field acquisition conditions

The durian disease dataset in (Hoang et al., 22 Jun 2026) contains 5,451 images across 10 classes and is described as essentially balanced, with class sizes ranging from 528 to 560. The source is a family-owned durian orchard plus four neighboring farms in Vinh Long, Vietnam. Images were acquired with an iPhone 14 under natural field conditions and natural lighting only; the reported challenges include motion blur, uneven illumination, diverse viewing angles, and complex or cluttered backgrounds.

Class	Images
Thrips disease	537
Stem blight	554
Canker disease	557
Pink disease	549
Stem cracking and gummosis	560
Mealybug infestation	536
Anthracnose	549
Fruit rot	528
Sooty mold	541
Yellow leaf disease	540

Unlike many leaf datasets, the durian set does not include a “healthy” class; all images belong to specific diseases or pest/infestation conditions (Hoang et al., 22 Jun 2026). Several classes refer to leaf infections and infestations, while others are more stem- or fruit-oriented but are captured in the same orchard context. This matters methodologically because performance gains on durian cannot be attributed to healthy-vs-diseased separation.

Image resolution ranges from 150 to 1824 pixels in width and 150 to 1452 pixels in height. For model input, all images are resized to $224 \times 224$ pixels. The split uses stratified sampling with 64% training, 16% validation, and 20% test, with random seed 42 for reproducibility. Durian-specific preprocessing and augmentation consist of color jitter with brightness $= 0.15$ and contrast $= 0.05$, random horizontal flip with $p = 0.5$, random rotation $(\pm 15^\circ)$, and Gaussian blur with kernel size $= 3$ and $\sigma \in [0.1, 1.0]$ (Hoang et al., 22 Jun 2026).

Because the class distribution is near-balanced, no class-weighting or focal loss is applied specifically for durian. The paper explicitly contrasts this with the potato benchmark, where severe imbalance is central. For durian, the reported benefit of the model is therefore interpreted primarily in terms of better feature integration and generalization rather than minority-class rescue (Hoang et al., 22 Jun 2026).

3. Cross-architectural Mixture-of-Experts for durian disease classification

The durian classifier in (Hoang et al., 22 Jun 2026) is not a durian-specific architecture but an adaptive soft Mixture-of-Experts that integrates EfficientNet-B0, DenseNet-121, and Swin-Tiny. Their stated complementarity is multi-scale representation, local texture modeling, and global or mid-range contextual modeling, respectively. The paper associates these strengths with disease manifestations such as small textured lesions, larger discoloration patterns, and broader leaf-, stem-, or fruit-level context.

The routing mechanism is input-dependent. Expert feature maps $F_{\text{Expert}_i}$ are projected into a shared $7 \times 7 \times 512$ space, pooled to 512-dimensional vectors, summed into a shared descriptor, and passed through a routing MLP with Linear $512 \rightarrow 256 \rightarrow 3$ followed by Softmax. The gating weights satisfy

$= 0.15$0

and the fused feature map is

$= 0.15$1

For durian, the classification head is GAP on $= 0.15$2, followed by dropout and a linear layer $= 0.15$3 with cross-entropy loss (Hoang et al., 22 Jun 2026).

Training uses a two-stage refinement schedule. Phase 1 fine-tunes for 100 epochs with batch size 32, Adam, learning rate $= 0.15$4, dropout 0.1, and early stopping patience 20. Phase 2 resumes from the best Phase 1 checkpoint for 20 epochs with batch size 32, learning rate $= 0.15$5, dropout 0.1, and patience 5 (Hoang et al., 22 Jun 2026). The paper attributes this refinement stage to improved generalization and stabilization.

Model	Accuracy (%)	F1-score (%)
MobileNet-V2	84.57	84.47
EfficientNet-B0	88.89	88.79
DenseNet-121	89.20	89.19
ResNet-50	88.73	89.25
Swin-Tiny	89.35	89.28
MoE (proposed)	93.98	94.03

The proposed MoE achieves 93.98% accuracy, 94.18% precision, 94.02% recall, and 94.03% F1-score on the durian test set (Hoang et al., 22 Jun 2026). Relative to Swin-Tiny, described as the best baseline in the paper’s durian comparison, the improvements are +4.63% accuracy, +4.31% precision, +4.58% recall, and +4.75% F1-score. The paper does not provide per-class metrics, confusion matrices, or durian-specific Grad-CAM visualizations, so any attribution of expert specialization for particular durian diseases remains inferential rather than directly demonstrated.

The significance of the durian result is methodological rather than botanical. Durian is used as an additional evaluation showing that the same fixed MoE design, selected and analyzed primarily on potato, can still obtain a 94.03% F1-score on a different crop and disease set collected under real orchard conditions (Hoang et al., 22 Jun 2026).

4. Durian plantation protection through hybrid YOLOv5-SSD and IoT

The orchard-protection system in (Shahrir et al., 2 Nov 2025) addresses animal intrusions in durian plantations, especially in Malaysia, where durian is described as a high-value export crop. The motivating animals are elephants, boars, and monkeys. Elephants are said to enter at night or dusk, trample trees, break branches, and eat foliage and fruits; the paper cites a case of approximately MYR 200,000 loss from one elephant incident. Boars uproot young trees and disturb roots and irrigation, while monkeys pick and eat unripe and ripe durian, break small branches, and cause scattered fruit drop.

The proposed architecture is a Raspberry Pi 5–based edge node with a webcam and an external speaker. Two detection models are run on the same video frame: YOLOv5 in PyTorch .pt format and SSD MobileNet in TensorFlow .tflite format. Detections are filtered by class $= 0.15$6 and confidence threshold

$= 0.15$7

The fusion rule is a logical OR ensemble: if either model detects an animal with confidence at least 0.5, the animal is considered detected (Shahrir et al., 2 Nov 2025).

Operationally, detection triggers a Telegram alert containing the detected class, confidence, and an image. The farmer can reply deter to activate a tiger-roar sound on loop through the external speaker, and stop to halt playback. The system is therefore a closed loop of detection, notification, deterrence, and remote supervision (Shahrir et al., 2 Nov 2025).

Training uses a small dataset of approximately 60 training images per class, with 10 test images per class for initial evaluation and an additional 20 tests per class for the confusion matrix. Annotation is performed with MakeSense.ai, and training is conducted on Google Colab with epoch counts of 50, 150, and 300 explored for mAP improvement (Shahrir et al., 2 Nov 2025). The paper reports the standard detection metrics

$= 0.15$8

with $= 0.15$9 as the area under the precision-recall curve and $= 0.05$0 as the mean of classwise APs.

Class	Accuracy	F1-score
Boar	0.85	0.92
Elephant	0.90	0.95
Monkey	0.70	0.82

The hybrid model achieves per-species real testing accuracies of elephant 90%, boar 85%, and monkey 70% (Shahrir et al., 2 Nov 2025). Monkeys are identified as harder to detect, plausibly because of smaller size, agile movements, and partial occlusions; the paper states this directly in the error analysis. Inference time on Raspberry Pi is prototype-level rather than video-frame-rate: SSD is reported at 12.7 s for boar and 9.3 s for elephant, YOLOv5 at 27.8 s for boar and 26.3 s for elephant, and the hybrid at 20.8 s for boar and 24.3 s for elephant (Shahrir et al., 2 Nov 2025). The authors accordingly frame the system as early warning rather than high-FPS surveillance.

The paper also reports a daytime advantage over nighttime conditions. Accuracy drops at night for elephants and boars, with poor lighting, shadows, and vegetation occlusions degrading performance. Suggested remedies include infrared or low-light cameras, additional illumination, more night training data, and potentially thermal cameras (Shahrir et al., 2 Nov 2025).

5. “Durian” as difficulty-aware group normalization in multimodal RLVR

In (Li et al., 25 Feb 2026), “Durian” does not denote the crop but the method name “difficulty-aware group normalization.” It is proposed for RLVR and GRPO in multimodal LLMs, where per-input std-based normalization of binary rollout rewards is described as unstable. Under vanilla GRPO, a rollout group uses

$= 0.05$1

with $= 0.05$2. The paper argues that multimodal settings frequently produce extreme groups—almost all 0s or all 1s—so the denominator becomes very small and advantage magnitudes become disproportionately large (Li et al., 25 Feb 2026).

Durian keeps the core RL objective but changes how std is estimated. It defines perceptual difficulty through visual entropy of patch features from the model’s visual encoder and reasoning difficulty through model confidence derived from rollout sequence log-probabilities. For image $= 0.05$3, the visual entropy is

$= 0.05$4

where $= 0.05$5 are eigenvalues of the patch covariance matrix. For reasoning difficulty, sequence log-probabilities are averaged across $= 0.05$6 rollouts: $= 0.05$7 Samples are regrouped by entropy quantiles for perceptual groups and by confidence quantiles for reasoning groups (Li et al., 25 Feb 2026).

The key change is that the numerator still uses the per-sample rollout mean, but the denominator uses std over a larger reward pool drawn from samples of similar difficulty. The perceptual advantage is

$= 0.05$8

and the reasoning advantage is defined analogously with $= 0.05$9. Durian then combines these with the original GRPO advantage: $p = 0.5$0 The paper emphasizes that this preserves intra-group reward distinctions while reducing sensitivity to extreme cases (Li et al., 25 Feb 2026).

The method is implemented atop Qwen2.5-VL-7B-Instruct with EasyR1 and evaluated on MathVerse, MathVision, MathVista, WeMath, and HallusionBench. Starting from the base model average of 53.3% over the five benchmarks, vanilla GRPO on 2.1K Geometry3K reaches 56.9%, while Durian on top of GRPO reaches 59.2%. With DAPO as backbone, vanilla DAPO reaches 57.3%, Durian on 2.1K reaches 59.3%, and Durian on ViRL39K reaches 60.1%, described as best overall (Li et al., 25 Feb 2026). The paper further states an average 11.3% improvement over the base Qwen2.5-VL across benchmarks.

The juxtaposition of crop-oriented durian studies with the algorithmic acronym “Durian” is terminologically notable. In recent arXiv usage, the term is therefore polysemous: one sense is agronomic and deployment-oriented, the other is an optimization method for multimodal reasoning (Li et al., 25 Feb 2026).

6. Limitations, open problems, and broader significance

The durian disease-classification study identifies several constraints on real deployment. The MoE architecture used for durian has 12.73 GFLOPs, 40.19 million parameters, and approximately 40 ms per image on Tesla T4, making it heavier than single backbones and potentially challenging for resource-limited edge devices (Hoang et al., 22 Jun 2026). The paper suggests model compression, distillation, or cloud offloading as possible remedies. It also notes that all durian images come from one region, Vinh Long, and from a small set of orchards, so broader robustness would require additional data from other regions, cultivars, growth stages, and seasons. A further limitation is the absence of a healthy class in the durian dataset, which restricts direct applicability to field settings where farmers also need healthy-vs-diseased discrimination (Hoang et al., 22 Jun 2026).

The orchard-protection system faces a different set of constraints. The dataset is small, nighttime performance degrades, inference on Raspberry Pi remains slow, multiple nodes would be required for large plantations, and the deterrent effectiveness of tiger-roar playback is not quantified by longitudinal field studies (Shahrir et al., 2 Nov 2025). The proposed future directions include more advanced detectors such as YOLOv7, YOLOv8, or transformer-based models if edge-efficient, model quantization and optimization, larger multi-orchard datasets, additional sensors such as thermal, motion, or audio, smarter automated deterrence, and multi-camera networks with centralized dashboards (Shahrir et al., 2 Nov 2025).

The RL method Durian also introduces trade-offs. It adds hyperparameters such as the number of reasoning groups and the coefficients $p = 0.5$1, requires computation of patch covariance, eigenvalue entropy, and token-level log-probabilities, and relies on difficulty proxies that may not perfectly match true conceptual difficulty (Li et al., 25 Feb 2026). The paper accordingly points to more precise difficulty estimation and adaptive grouping strategies as future work.

Taken together, these studies show that durian has become a nontrivial reference point in current machine-learning research. As a crop, it anchors work on disease recognition and orchard protection under realistic sensing constraints. As an acronym, it names a stabilization method for multimodal RLVR. The common thread is not taxonomy but the design of robust inference and optimization procedures under heterogeneous, high-variance conditions (Hoang et al., 22 Jun 2026, Shahrir et al., 2 Nov 2025, Li et al., 25 Feb 2026).