Ammar Aslan

Architectures for building extraction are usually compared under different training and inference set-ups, which makes the reported performance figures difficult to compare directly. This study evaluates three convolutional segmentation architectures (U-Net, U-Net++, DeepLabV3+) and one transformer-based architecture (SegFormer) on the aerial orthophoto subset of the WHU Building Dataset under a single training and inference protocol. Fixing the encoder to ResNet-34 in the convolutional models separates the effect of decoder design from backbone capacity. U Net++ achieved the highest performance at 89.04% IoU and 94.20% F1, and the difference between the models remained within a narrow range of 2.42 points on the standard IoU metric. On the boundary IoU metric, which measures overlap only within the edge band, the same spread rose to 9.26 points, roughly 3.8 times as wide. Reporting standard IoU alone therefore systematically understates the difference between architectures. An inference protocol combining an exponential moving average, eight-way test-time augmentation and a decision threshold tuned on the validation set yielded gains of 0.47 to 0.69 IoU points without changing the ranking of the models; that magnitude corresponds to roughly a quarter of the total spread attributable to architecture. On the cost side, parameter count proved a misleading indicator: a 6.7% difference in parameters between U-Net++ and U-Net corresponded to a 135.2% difference in multiply accumulate operations. The results show that model selection in building extraction cannot rest on a single accuracy figure, and that boundary-based metrics should govern the choice in applications where the position of the boundary is critical.

Achieving sustainable development goals and mitigating the effects of climate change necessitates meeting global energy demand from clean sources. In this context, wind energy emerges as one of the most promising sources. Due to the variable and intermittent nature of wind, accurately predicting the power generated from wind energy systems is crucial for integrating these sources into the grid. While current advanced prediction models can provide this accuracy, disadvantages such as high computational complexity, large memory footprint, and low inference speeds seriously limit the practical applicability of these models in real-time and resource-constrained environments. This study proposes a new and lightweight architecture, the Intelligent Feature-Engineered Minimal Multi Layer Perceptron (IFE-MLP) model, as a feature engineering-based alternative to complex transfer learning approaches. Unlike classical transfer learning paradigms that rely on parameter transfer or domain adaptation, the proposed model achieves competitive prediction accuracy through intelligent feature construction while maintaining high computational efficiency. The performance of the proposed IFE-MLP model has been comprehensively compared with two transfer learning approaches: Fine-Tuning-Based Transfer Learning (FTTL) and Feature-Based Transfer Learning (FBTL). Data obtained from a wind power plant in Bursa, Turkey, was used for performance evaluation. The prediction results revealed that the proposed model exhibited only slightly lower prediction accuracy compared to the FTTL model. However, the main advantage of the proposed IFE-MLP model is clearly evident in terms of computational efficiency and resource utilization. The proposed model achieved its prediction performance using 450 times less memory compared to the FTTL model and 2,050 times less memory compared to the FBTL model. Furthermore, it reduced the training time by 2.2 times by requiring 546 times fewer training Floating Point Operations (FLOPs) and 435 times fewer inference FLOPs, respectively. Inference latency, a critical metric for real-time applications, was measured at 0.05-0.08 ms, corresponding to a 16-100x improvement over the compared models. The conducted decomposition study confirmed that the intelligent feature engineering components in the model architecture provided incremental and meaningful contributions to performance. The findings indicate that the proposed IFE-MLP architecture offers comparable prediction accuracy alongside a significant improvement in resource usage. These characteristics suggest that the model could be a suitable alternative for resource-constrained environments and real-time applications.

Short-term wind power forecasting supports grid stability despite wind’s intermittency, but deep learning (DL) models often require excessive computation, hindering real-time deployment. This study compares three classical regression models, linear regression (LR), k-nearest neighbors (KNNs), and support vector regression (SVR) using 10-minute SCADA data. The data retained raw variability, with only negative power values set to zero. Hyperparameters of KNN and SVR were optimized using GridSearchCV with time series cross-validation. Optimized SVR achieved superior performance (MAE = 183.93 kW, RMSE = 387.64 kW, R ² = 0.91, MAPE = 144.88%, MdAE = 84.44 kW), outperforming KNN (R ² = 0.79) and LR (R ² = 0.91 but higher errors). Scatter plots and residuals highlighted SVR’s ability to handle nonlinearities. Results suggest that SVR is a viable candidate SVR for low-resource, real-time forecasting in renewables.

Brain tumors can cause serious neurological damage and death by putting pressure on critical brain regions that manage vital functions. Given the complex structures in the brain, human error in the evaluation of radiological images can create difficulties in the detection of these tumors. Convolutional neural networks (CNNs) are type of deep learning (DL) and are widely used, especially for analyzing visual data. The advantage of CNNs in detecting brain tumors is that they can automatically learn features from images and minimize human error by increasing the classification accuracy. In this study, a unique CNN-based model is proposed for brain tumor diagnosis using magnetic resonance imaging (MRI) images. A high classification score was obtained using a dataset consisting of 3096 MRI images divided into four categories: glioma, meningioma, normal brain, and pituitary tumor. The model achieved an overall 93% accuracy rate in tumor detection. In particular, great success was seen for the detection of pituitary tumors with 96% precision and a 95% F1 score. This study demonstrates that DL has significant potential in medical image analysis. The novelty of our approach lies in designing a lightweight CNN architecture from scratch that achieves high accuracy without relying on transfer learning, while requiring significantly fewer computational resources than traditional deep architectures.