Deep learning architectures are widely used in state-of-the-art image classification tasks. Deep learning has enhanced the ability to automatically detect and classify plant diseases. However, in practice, disease classification problems are treated as black-box methods. Thus, it is difficult to trust the model that it truly identifies the region of the disease in the image; it may simply use unrelated surroundings for classification. Visualization techniques can help determine important areas for the model by highlighting the region responsible for the classification. In this study, we present a methodology for visualizing coffee diseases using different visualization approaches. Our goal is to visualize aspects of a coffee disease to obtain insight into what the model “sees” as it learns to classify healthy and non-healthy images. In addition, visualization helped us identify misclassifications and led us to propose a guided approach for coffee disease classification. The guided approach achieved a classification accuracy of 98% compared to the 77% of naïve approach on the Robusta coffee leaf image dataset. The visualization methods considered in this study were Grad-CAM, Grad-CAM++, and Score-CAM. We also provided a visual comparison of the visualization methods.
Deep learning-based methods have become an active research area in medical imaging. Malaria is diagnosed by testing red blood cells. Deep learning methods can be used to distinguish malaria infected cell images from non-infected cell images. The small number of malaria dataset may limit the application of deep learning. Moreover, the infected area in the cell images is generally vague and small, requiring more complex models and a larger dataset to train on. Motivated by the tendency of humans to highlight important words when reading, we propose a simple neural network training strategy for highlighting the infected pixel regions that are mainly responsible for malaria cell classification. In our experiments on the NIH(National Institutes of Health) malaria dataset available in public domain, the proposed method significantly improved classification accuracy for our four different sized models, ranging from simple to complex including Resnet and Mobilenet. Our proposed method significantly improved classification accuracy. The result indicate that approach achieves a classification accuracy of 97.2%, compared to 94.49% for a baseline model. In addition, we demonstrate the superiority of the proposed strategy by providing an analysis on the magnitude of weight parameters in terms of regularization.
27 February 2025 in IEEE Transactions on Geoscience and Remote Sensing
Analyzing radargrams obtained from radar sounder instruments is an effective method for studying the subsurface of celestial bodies. However, existing deep learning-based methods for automatic classification of subsurface in radar sounder data require a large amount of labeled training data, posing a significant challenge. Moreover, these methods lack generalization to radargrams from different campaigns due to varying distributions within the same classes. To address these limitations, we propose a novel few-shot pixel-based classification framework for radar sounder data. This framework aims to learn underlying patterns using only a few labeled support samples and adapt quickly to radargrams from different campaigns with minimal labeled information without requiring retraining. Given the scarcity of labeled radar sounder data and the significance of low-level features (like texture and intensity) in class differentiation over high-level features (like shapes), we propose simplifying the segmentation task into a classification problem by treating each pixel independently. To preserve spatial information, we introduce spatial input by integrating neighboring pixels in the depth dimension and incorporating sequence awareness to address misinterpretations of reflections from different subsurfaces. Furthermore, unlike conventional semantic segmentation approaches relying on encoder-decoder structures, our framework eliminates the need for the decoder component. We evaluate the proposed method on different datasets acquired in regions of Antarctica by MCoRDS. Results show the effectiveness and generalization capability of the proposed framework to accurately segment different subsurface targets under very limited data.
18 December 2025 in IEEE Geoscience and Remote Sensing Letters
Automatic semantic segmentation of radar sounder data is a critical task for analyzing subsurface structures in planetary and terrestrial studies. In this context, one of the major issues for the development of deep learning models for semantic segmentation is the lack of labeled data, which affects the training of neural architectures. In this study, we propose a semi-supervised learning method that leverages scribble annotations, including points, diagonal lines, and polygons, alongside pseudo-labels generated from unlabeled data. To ensure the reliability of pseudo-labels, we propose selecting them based on confidence and spatial proximity constraints. To validate our method, we performed extensive experiments on terrestrial and planetary datasets. The results demonstrate that our approach consistently outperforms existing semi-supervised methods when trained on scribble annotations. In particular, the proposed method trained on diagonal scribbles yields the highest overall accuracy of 99.4% on the terrestrial dataset and 94.7% on the planetary dataset. These findings indicate that the proposed method achieves performance comparable to those of the fully supervised methods trained on dense annotations, while significantly reducing labeling costs.