Using trained deep learning models on an out-of-distribution dataset could affect those models’ performance, which is most likely to decline due to the domain shift between the training domain and the target domain. To this end, Domain Generalization (DG) is one of the tools used to improve the generalization rate of the model and increase its adaptability to unseen data. Using a data-based domain generalization, this study intends to create a solar module segmentation model using thermal UAV images, by including various type of technologies and representations of solar modules in training data. Developing the generalization of the solar panel segmentation could enhance the performance in the workflow of automatic photovoltaic plants monitoring. The thermal images dataset was collected from six photovoltaic sites in Benguerir - Morocco in various shooting conditions. The global capacity of these systems is 1MW and they are equipped with different technologies. The training of this model achieved 99% [email protected] for testing and could perform well on unseen data, that is different from the training data, ensuring the effectiveness of the approach to generalize the trained Deep learning models which can be used in other domains.

One of the foremost answers to the challenge of food security lies in precise prediction of crop yields. This prediction task is complex due to the numerous variables involved. A primary source for yield prediction is historical remote sensing data, notably satellite imagery. However, such imagery often includes redundant data because of its multiple spectral bands and the extensive archived historical data over extended periods. In this study, we introduce an advanced framework for corn yield prediction at field level utilizing Long Short-Term Memory (LSTM) combined with attention mechanisms and randomized Principal Component Analysis (r-PCA). This framework aims to filter out redundant data and improve yield prediction using raw Sentinel-2 imagery, and incorporating stacked temporal and spectral data. Our model was tested using data of the year 2019, and was compared against the simple LSTM architecture using key performance indicators. Our proposed framework achieved a Mean Absolute Error (MAE) of 1.40, surpassing the MAE recorded by the LSTM. The results validate the effectiveness of our approach, particularly for field-scale predictions, demonstrating the enhanced capability of our model in accurately forecasting crop yields at this scale.

Groundwater withdrawals by farmers, in Morocco, are very numerous and informal. Therefore, the need for information on the location of irrigated areas is becoming increasingly important. Our main objective, in this study, is to evaluate the use of high-resolution Landsat 8 (L8) time series images and Random forest (RF) method to produce a land cover map with a sufficient precision to monitor the extension of irrigated areas. In the first part of this study, four parameters were evaluated: Number of trees, min split samples, max features and max depth. The results proves that the last parameter is the most important and has more impact on the oob score, which can reach 91 %. The second part of this study was devoted to reduce furthermore the number of features taken as input in the classification process. This was done through feature reduction then selection. The computational time was highly reduced and the best level of classification accuracy was reached by using only Landsat 8 (L8) time series images, statistics on the temporal spectral indices (NDVI, MNDWI) and Range texture.

The goal of our research is to develop a new objectoriented approach of image analysis for extraction of urban objects from very high spatial resolution images. The proposed approach is constituted of two stages: a passage of pixels to primitive objects and a passage of primitives to final objects. The first stage leaves from pixels to create primitives of objects by using a new multispectral non parametric segmentation approach. For the second stage, it is based on a fuzzy rule base and a spatial analysis. The proposed approach was applied on an Ikonos image of Sherbrooke (Canada). A confrontation with the ground truth gave a rate of 85% of good detection. These results are encouraging because the new approach does not requires any prior information or training data.