Simrandeep Singh

Accurate prediction of crop yields is fundamental to achieving food security, enhancing farm management, and making data-driven decisions due to changing climates. Conventional statistical and machine learning techniques typically do not adequately capture the complex ways that different sources of agricultural data interact over both space and time. Therefore, this research introduces MultiYieldFormer, a novel transformer-based method to forecast crop yield using a wide variety of time-series data collected from remote sensing and environmental factors in combination. MultiYieldFormer is a framework that integrates Sentinel-1 SAR observation data, Sentinel-2 multispectral imagery, climate data (including temperature and rainfall, soil moisture levels), and seasonal growth data to improve the modeling of how crops grow and change over time. It uses self-attention and transformer network structures to capture how crops grow and change from one time to another as well as how these changes are related across different types of features. By combining multi-type feature information into seasonal phenology patterns, it allows users to gain insights from the seasonal fluctuations in their crop yields. Using comprehensive multi-year crop datasets representing Kharif crops including rice, maize, and moong in the Ludhiana area of Punjab state in India, the results of this experiment have shown that the MultiYieldFormer model improves prediction performance by capturing effective nonlinear relationships between weather and environmental conditions through satellite observations. The proposed MultiYieldFormer achieved an average value for coefficient of determination (R²) of 0.89, Root Mean Square Error (RMSE) value of 0.42 t/ha, Mean Absolute Error (MAE) value of 0.33 t/ha, and Mean Absolute Percentage Error (MAPE) value of 5.61%. The results of R² were recorded as 0.91 for rice, 0.88 for maize, and 0.89 for moong crop. Comparisons of performance suggest that the model outperforms traditional machine learning (ML) and deep learning (DL) approaches in terms of generalization and robustness. This proposed framework provides a scalable and transferable approach to estimating regional yield from crops and can serve as a tool for precision agriculture applications, agricultural planning, and sustainable management of crops.

Skin cancer continues to be one of the most frequently diagnosed malignancies worldwide, with early diagnosis being vital for improving patient survival and allowing for timely treatment. In this work, we proposed the adaptive fusion-based deep ensemble framework for multi-class skin lesion classification. It integrates four state-of-the-art pretrained CNNs, namely Xception, EfficientNetB4, InceptionResNetV2, and MobileNetV2, to exploit their complementary feature extraction capabilities and enhance lesion representation. Unlike conventional ensemble approaches that employ static weighting strategies, the proposed framework employs an instance-specific adaptive weighting mechanism. It dynamically combines the output from these individual models based on prediction confidence, entropy-based uncertainty, and model reliability. This adaptive fusion strategy enables the ensemble to assign optimal weights to the most reliable model(s) for each input image, thereby improving robustness and classification performance. The proposed framework obtained a 98.86% test accuracy, 98.70% validation accuracy, 98.88% weighted precision, 98.86% weighted recall, and 98.85% weighted F1-score on the HAM10000 database.

The conversion of fruit waste (FW) into biofertilizers and biopesticides, thereby offering a way of managing agricultural and food industry waste and at the same time contributing to green agriculture. Fruits and their wastes contain nutrients, organic matters, and bioactive compounds that are important for the production of microbial biofertilizers and natural pest repellents. This transformation is not only a way to redeem the fertility of the soil and health of the plants but also a way to decrease the intensive usage of chemicals that have negative influence on the environment and human health. Some of the techniques used include composting, anaerobic digestion, fermentation, and green extraction techniques to recycle the above wastes into agriculturally useful products. More recent incorporations suggest the ideal use of bioformulations produced from FW in farming systems, backed up by concepts on cost implications and the possibility of scale-up in industrial-level examples. The economic feasibility or advantage of this approach lies in the low cost of raw material and the added value from waste product. Though some concerns are still there, such as fluctuation in the waste generation frequency throughout the year, the lack of standardization, the limitation of the laws, and problems arising from the microbial stability and shelf life of the final product. To overcome all these limitations, future directions are given toward high biotechnological events, advanced implements of precision agriculture, and generation of combined multipurpose formulations of biofertilizers and biopesticides. To improve the commercialization and increase the societal acceptance of such products, the use of circular economy principles and increased collaboration between government and companies is encouraged. In conclusion, utilizing FW as biofertilizers and biopesticides remains a strong and sustainable solution for regenerative agriculture.

Deep learning (DL) has significantly transformed computer vision, demonstrating remarkable achievements and extensive real‐world applications. However, recent studies have highlighted a critical vulnerability of DL models to adversarial examples (AE), where slight perturbations in input data can lead to erroneous outputs. We observe that the behaviour of the AE is similar to a chaotic system, where a minor change in the input leads to a significantly different output. In response, we propose a novel approach for detecting and categorizing adversarial inputs encountered by classification neural networks. The proposed approach focuses on extracting statistical profiles, termed as chaotic feature vectors (CFVs), from a collection of features derived from ordinal patterns (OP). In this work, the proposed AE detection method is tested on seven attack methods and three image datasets including MNIST, FMNIST and CIFAR10. The results indicate that CFVs exhibit promising capabilities in discerning AE against various types of adversarial attacks on different datasets. This advancement lays the foundation for devising attack mitigation strategies, thereby enhancing the robustness and security of DL models in the face of adversarial threats.

American scientific publishing group

Image segmentation is a critical stage in the analysis and pre-processing of images. It comprises dividing the pixels according to threshold values into several segments depending on their intensity levels. Selecting the best threshold values is the most challenging task in segmentation. Because of their simplicity, resilience, reduced convergence time, and accuracy, standard multi-level thresholding (MT) approaches are more effective than bi-level thresholding methods. With increasing thresholds, computer complexity grows exponentially. A considerable number of metaheuristics were used to optimize these problems. One of the best image segmentation methods is Otsu’s between-class variance. It maximizes the between-class variance to determine image threshold values. In this manuscript, a new modified Otsu function is proposed that hybridizes the concept of Otsu’s between class variance and Kapur’s entropy. For Kapur’s entropy, a threshold value of an image is selected by maximizing the entropy of the object and background pixels. The proposed modified Otsu technique combines the ability to find an optimal threshold that maximizes the overall entropy from Kapur’s and the maximum variance value of the different classes from Otsu. The novelty of the proposal is the merging of two methodologies. Clearly, Otsu’s variance could be improved since the entropy (Kapur) is a method used to verify the uncertainty of a set of information. This paper applies the proposed technique over a set of images with diverse histograms, which are taken from Berkeley Segmentation Data Set 500 (BSDS500). For the search capability of the segmentation methodology, the Arithmetic Optimization algorithm (AOA), the Hybrid Dragonfly algorithm, and Firefly Algorithm (HDAFA) are employed. The proposed approach is compared with the existing state-of-art objective function of Otsu and Kapur. Qualitative experimental outcomes demonstrate that modified Otsu is highly efficient in terms of performance metrics such as PSNR, mean, threshold values, number of iterations taken to converge, and image segmentation quality.