The Bodoland Territorial Region, an autonomous sixth schedule area in Assam, North East India, is a botanical treasure trove. Forests and designated reserve forests cover roughly 30 to 35 percent of its total geographical area, and the region hosts three national parks. For the predominantly Bodo and other indigenous communities who inhabit its five districts, wild edible vegetables are not exotic curiosities but staples of the regular diet, prized for their nutritional and therapeutic properties. These plants underpin food security and socio-economic development in rural communities, yet they are gradually losing popularity among the younger generation, who often lack the knowledge to recognize them. Manual identification is time-consuming and prone to error, creating a pressing need for an automatic system that can help anyone identify these plants confidently.

Existing approaches to plant identification, whether based on classical machine learning algorithms or deep convolutional neural networks, excel when images are captured under controlled or pristine conditions. But the real world of forests and grasslands is unforgiving. Complex natural backgrounds, variable lighting, partial occlusions, and cluttered vegetation conspire to degrade the performance of even the most sophisticated pre-trained models. Moreover, the computational requirements of these architectures present a formidable barrier to deployment on low-resource edge devices such as smartphones, agricultural drones, and embedded systems. The research team, led by Tania Sarkar and Pranav Kumar Singh of the Central Institute of Technology Kokrajhar, alongside Parag Bhuyan and Anshuman Kalla, set out to bridge this gap with a novel architecture that adaptively adjusts its computation according to the complexity and information content of the input.

To train and evaluate their model, the researchers assembled a dataset of 2310 images spanning 21 wild edible vegetable species, with 110 images per class. The images were collected across various districts of BTR, primarily from village marketplaces and surrounding forest areas, using ordinary camera-equipped mobile phones. Plants were photographed against their natural backgrounds from multiple angles under different lighting conditions, at different growth stages and seasons, ensuring the dataset captured the full spectrum of real-world variability. Local villagers and vendors were consulted to verify correct identification, and cross-verification relied on published literature as well as preserved specimens from the District Agriculture Office in Kokrajhar, the Department of Botany at Bodoland University, and Gauhati University. The dataset was split into 80 percent training and 20 percent testing, with a validation set drawn from the training data to prevent overfitting.

The architectural innovation of DynaLiteNet lies in its coordinated integration of four adaptive modules. The Dynamic Kernel Selection block employs three parallel convolution branches with kernel sizes of 1×1, 3×3, and 5×5, capturing fine-grained texture details, intermediate spatial structures, and coarse regions respectively. Rather than treating these branches uniformly, the module computes the local entropy of the feature map—a measure of information variability—and uses it to assign adaptive attention weights to each branch through a softmax-like formulation. This allows the model to engage larger receptive fields in complex areas while suppressing superfluous calculations in simpler ones, a strategy that directly contrasts with traditional CNNs where convolutional kernels remain fixed throughout inference.

Complementing DKS, the Region-wise Adaptive Feature Routing module introduces spatial conditional computation. The feature map is partitioned into N×N regions, and an importance score for each region is derived from regional entropy and mean activation statistics. Regions scoring above a routing threshold proceed through the adaptive feature-processing stream, while low-information regions are passed through identity mappings, effectively bypassing expensive convolutional operations. Because plant symptoms typically occupy only localized areas of an image, with the remainder dominated by background, this selective routing dramatically reduces redundant computation. Unlike transformer attention mechanisms that incur quadratic computational costs, RAFR achieves dynamic spatial adaptivity without any overhead.

The third module, Ultra-Light Context Encoding, addresses the need for global contextual awareness without the burden of self-attention. Inspired by context encoding algorithms, ULCE computes a global feature descriptor by averaging spatial responses and then modulates the feature map through a learnable scaling parameter, allowing the network to capture spatial co-occurrence patterns and lesion distribution with a small number of trainable parameters. Finally, Entropy-Guided Global Pooling replaces conventional Global Average Pooling, which treats every feature location equally. Instead, EGGP computes an entropy-derived weight map that adaptively modifies the contribution of each spatial response during global aggregation, ensuring that informative regions play a dominant role in the final feature representation while uniform or low-information areas are downweighted.

Experimental results demonstrate the potency of this coordinated design. DynaLiteNet achieves a classification accuracy of 90.69 percent across the 21 species, evaluated over ten different random splits with early stopping at 50 epochs. The model was trained using the AdamW optimizer with a learning rate of 0.001, which provided better regularization and convergence stability than Adagrad or Stochastic Gradient Descent. In comparative benchmarks, DynaLiteNet outperformed popular pre-trained architectures including AlexNet, ResNet50, DenseNet121, and InceptionV3, as well as efficient models such as MobileNetV3, ShuffleNetV2, EfficientNet-Lite, and MobileViT. Crucially, the architecture accomplishes this with merely 16.5 thousand trainable parameters and 0.18 GFLOPs, a computational profile orders of magnitude smaller than conventional deep CNNs, making it exceptionally well-suited for real-time precision agriculture and low-resource edge deployment.

An extensive ablation study validated the contribution of each module by progressively adding them to a base network under uniform hyperparameter configurations. The entropy-guided modules enhanced important lesion areas and suppressed useless background features, facilitating localization. Dynamic Kernel Selection improved multi-scale feature capture, Region-wise Adaptive Feature Routing reduced computational overhead by skipping low-information areas, Ultra-Light Context Encoding captured long-range spatial links with minimal parameters, and Entropy-Guided Global Pooling improved aggregation of discriminative features by focusing on sensitive areas. A five-fold stratified cross-validation confirmed stable performance, with average accuracy of 90.69 percent and a standard deviation of only 0.40 percent, alongside minimal fluctuation in precision, recall, and F1-score across folds.

Interpretability was assessed through Grad-CAM visualizations, which revealed that the model primarily attends to discriminative regions of vegetable leaves—unique shape, texture, venation, and morphological features—when classifying correctly. Misclassified samples, conversely, exhibited widespread activation or focus on visually confusing areas, with errors most prominent in data exhibiting high inter-class similarity, background noise, lighting changes, partial occlusion, or insufficiently unique physical characteristics. The researchers acknowledge limitations, including the relatively small dataset size, geographic focus on BTR, and the absence of direct implementation on dedicated edge hardware. Future work will explore dataset expansion, multi-region validation, deployment on Raspberry Pi and NVIDIA Jetson platforms, and the development of a mobile application to bring this technology directly to the rural communities of BTR who depend on these wild edible vegetables for their livelihood and wellbeing.

Subject of Research: An entropy-guided dynamic lightweight network for identification of wild edible vegetables of BTR

Article Title: An entropy-guided dynamic lightweight network for identification of wild edible vegetables of BTR

Article References: Sarkar, T., Singh, P. K., Bhuyan, P., & Kalla, A. (2026). An entropy-guided dynamic lightweight network for identification of wild edible vegetables of BTR. Discover Artificial Intelligence, 6(1), Article 1539. https://doi.org/10.1007/s44163-026-02528-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02528-x

Keywords: entropy-guided, dynamic, lightweight, network, identification, wild, edible, vegetables, scientific research