Research & Development

Abstract Synopsis

Estimating carbon levels across land and ecosystems is gaining momentum as climate policies, sustainable practices, and carbon credit programs have become major talking points in recent years. Carbon credits are hugely beneficial assets for the environment, businesses, industries and individuals. So far, existing methods are heavily dependent on fieldwork and complex models, making the application complex and difficult to scale. Despite increasing availability of satellite data, there remains a clear research gap in developing scalable, lightweight frameworks that enable near real-time carbon estimation using only remote sensing, without dependence on ground measurements. To address this gap, this paper puts forward a cloud based platform that utilizes machine learning for effective real-time assessment of Above Ground Biomass (AGB), Below Ground Biomass (BGB), and Soil Organic Carbon (SOC) over a specified plot of land. The system enables time-based comparison across different years, making it suitable for monitoring carbon change and supporting preliminary carbon credit evaluation. The performance of the proposed models is demonstrated using standard regression metrics. The AGB model achieved an R2 of 0.73 with an RMSE of 46.11 tonnes per hectare, while the SOC model achieved an R2 of 0.60 with an RMSE of 1.69 g/kg. These results indicate that the proposed framework provides reliable and scalable carbon estimates, suitable for rapid assessment and decision making support.

Abstract Synopsis

Wild edible plants play a significant role in the socio-economic conditions of various communities by enhancing food security, providing economic benefits and supporting sustainable livelihoods. Proper identification of these plants is crucial in preserving cultural heritage and promoting environmental conservation. With the popularity of AI, automatic plant identification has gained much attention in the last decade. Machine learning, especially, deep learning has proven its efficiency in identifying these plants successfully. However, due to the requirement of massive amounts of data, transfer learning is often preferred over traditional deep learning algorithms. Our study demonstrates the use of transfer learning for the identification of wild edible vegetables in Bodoland Territorial Region (BTR). BTR, an autonomous region in Assam, Northeast India is extremely rich in plant resources. Pre-trained models Xception, VGG16 and Resnet50 are fine-tuned as well as used as feature extractors to develop 12 classification models. A new dataset is created comprising of 21 species of wild edible vegetables of BTR and is used for training our models. An in-depth analysis is done to investigate the performance of these models on the above classification task. By developing an automatic plant identification system, this study aims to shed light on the available plant resources of BTR and contribute to the enhancement of the quality of life of its rural communities.

Abstract Synopsis

Maize is a crucial crop globally, playing a significant role in economic systems. Timely detection of diseases in maize leaves is essential to ensure high crop yields. However, traditional manual inspection methods are slow and often inaccurate, highlighting the need for automated approaches. The integration of machine learning and IoT in smart farming offers promising solutions for early disease detection and prevention, which are key to sustainable agriculture. Deep learning has been increasingly utilized in identifying plant leaf diseases, with many systems employing vision-based machine learning for real-time detection. Convolutional Neural Networks (CNNs) have delivered impressive results in this area. However, the newer concept of Vision Transformers (ViTs) in vision-based deep learning is gaining attention. Although ViTs have shown potential in image classification, their application in plant leaf disease classification is still underexplored. This paper introduces MaizeViT, a hybrid model that combines Vision Transformer and CNN specifically for maize leaf disease classification. MaizeViT’s performance is evaluated using publicly available datasets, and it outperforms standard CNN models such as VGG16, DenseNet121, ResNet50, MobileNetV2, and InceptionV3. The results demonstrate that MaizeViT achieves an average accuracy of 98.15% and a precision of 98.31%.

Abstract Synopsis

Identifying and diagnosing plant diseases presents a considerable challenge. Recognition and prevention of crop diseases are important for maintaining healthy plant growth to ensure sustainable supply and food security for the world’s fast-increasing population. Manual examination of plant diseases is costly, slow, non-scalable, labor-intensive, and error-prone. Farmers have traditionally used manual methods to diagnose and classify plant leaf diseases, which can be imprecise and impracticable for large-scale applications. Farmers have the potential to minimize losses and enhance crop productivity through the application of automated image processing techniques. Researchers have established a variety of techniques to identify and classify plant leaf diseases by analyzing images of affected leaves. In this work, we analyze two attention mechanism-based models, SE_SPnet and Res4net-CBAM, and compare these two models with four standard CNN models, Resnet50, VGG16, DenseNet121, and InceptionV3. We evaluate these models using datasets of leaf diseases from three distinct plants: rice, tea, and maize. The results of the experiments show that Res4net-CBAM is the best model. It had an average accuracy of 99.78% on the rice leaf disease dataset, 98.27% on the tea leaf disease dataset, and 97.97% on the maize leaf disease dataset, which was better than the other models.

Abstract Synopsis

This survey presents a comprehensive overview of Legal Natural Language Processing (NLP) in the Indian context, with a focus on linguistic diversity across Indian languages and challenges related to equitable access to legal resources. Based on a decade-long analysis of law-centric literature, we trace the evolution of legal NLP research in India, highlighting the adoption of deep learning architectures, pre-trained language models, and domain-specific embeddings. We identify key application areas, such as legal named entity recognition, judgment prediction, legal question answering, case summarization and other key legal NLP tasks. The survey also emphasizes the need for open data sets and reproducible code to foster transparency and scalability in legal NLP research. This work aims to inform researchers, practitioners, and policymakers about current developments, challenges, and future directions in this rapidly evolving field.