This work aims to assess the effectiveness of machine learning (ML) algorithms and semi-empirical models for surface soil moisture (SSM) retrieval by exploring the Sentinel-1 backscatter and interferometric coherence data. Firstly, three commonly used categories of ML algorithms are evaluated using data gathered from diverse rainfed and irrigated wheat fields located in Morocco and Tunisia. Specifically, these algorithms include: artificial neural network (ANN), deep neural network (DNN), three support vector regression (SVR) models (radial basis function (SVR_rbf), linear (SVR_linear) and polynomial (SVR_quad) kernels) and two tree-based methods (Random Forest (RF) and XGBoost). The comparison between predicted and measured SSM showed that the best retrieval results were obtained using Sentinel-1 data at VV polarization with R ranging between $ 0.68$ and $ 0.76$ and RMSE of $ 0.05 m^{3}/m^{3}$ and $ 0.06 m^{3}/m^{3}$ . Secondly, to further assess their transferability, the ANN, SVR_rbf and XGBoost that demonstrated the most favorable results from each category were evaluated and compared against the coupled Water Cloud and Oh models (WCM), using a second dataset collected over a drip-irrigated wheat field in Morocco. Overall, the best retrieval results were achieved by ANN and SVR_rbf with R and RMSE of $ 0.81$ and $ 0.034 m^{3}/m^{3}$ , respectively. In addition, their performances were consistent with that of WCM which yielded R and RMSE values of $ 0.81$ and $ 0.04 m^{3}/m^{3}$ , respectively. Finally, due to its good compromise between retrieval accuracy of SSM, processing time and simplicity, SVR_rbf was chosen to generate high-resolution SSM maps from Sentinel-1 data over irrigated wheat fields.

We propose a deep neural network for detecting the severity levels of diabetic retinopathy disease by analyzing high-resolution retinography images. The architecture of this network is heuristically constructed starting with the simplest possible structure with no hidden layers, and progressively improving it by adding three types of hidden layers: fully connected, convolutional, and pooling layers. The training, validation, and generalization test of this architecture were performed on a real-world, open-access, and BSD-licensed dataset containing 3662 high-resolution color images, 72% of which were used to train the model while 20% were reserved for the validation process, and 8% for the generalization test on unseen images. The experimental results show that the proposed network yields excellent results in detecting the presence or the absence of the disease, and very good and promising results in distinguishing between its different levels of severity.

In this paper, we propose a new neural architecture for object classification, made up from a set of competitive layers whose number and size are dynamically learned from training data using a two-step process that combines unsupervised and supervised learning modes. The first step consists in finding a set of one or more optimal prototypes for each of the c classes that form the training data. For this, it uses the unsupervised learning and prototype generator algorithm called fuzzy learning vector quantization (FLVQ). The second step aims to assess the quality of the learned prototypes in terms of classification results. For this, the c classes are reconstructed by assigning each object to the class represented by its nearest prototype, and the obtained results are compared to the original classes. If one or more constructed classes differ from the original ones, the corresponding prototypes are not validated and the whole process is repeated for all misclassified objects, using additional competitive layers, until no difference persists between the constructed and the original classes or a maximum number of layers is reached. Experimental results show the effectiveness of the proposed method on a variety of well-known benchmark data sets.

We propose a competitive fuzzy-neuro model for image compression. This model is based on a new unsupervised fuzzy learning algorithm, designed for optimal training of competitive neural networks. Experimental results show that the proposed model can perform better than other well-known methods of its category, including FCM and IFLVQ. Typical examples of these results are presented and discussed.