Research

Current research

Explainability and signal processing

My first topic of interest is to improve our understanding of deep learning models. In computer vision, current methods focus on assessing where model focus on input images. However, these approaches are uninformative when it comes to explain why classification models can fail to correctly predict some instances.

My research consists in using tools from classical signal processins theory, in particular the wavelet transform, to meaningfully decompose the input image and then to identify the important parts of this decomposition into the model's prediction. I prefer using the wavelet transform over the Fourier transform as the wavelet decomposition of an image is localized in space and in frequency, which on images enables to assess where and what (scales) models see.

During the PhD thesis, we introduced the wavelet scale attribution method (WCAM), an extension of Fel's et. al. attribution method to the wavelet domain in order to assess what models see on an image. Scales of the wavelet transform correspond to structural components of the image: therefore, we can assess whether the model relies on shapes, textures or other intermediary components to make its prediction.

I believe that the the WCAM can be expanded into several interesting directions:

• First, it can be expanded to 1D signals such as sounds,
• Second, I believe that there can be interesting connections between the scale decomposition obtained with the WCAM and recent concept-based XAI approaches such as CRAFT.
• Finally, the scales of the wavelet transform are localized in frequency, and several works (see for instance this work or this one) leveraged frequency-based methods to characterize and analyze the robustness of models to various input perturbations.

Improving the observability of rooftop PV systems

The photovoltaic (PV) installed capacity grows very quickly, both worldwide and in France. A specificity of PV energy is that a sizezable share of the systems are located on rooftops (20% of the installed capacity in France). These rooftop systems are not well known by authorities or transmission system operators. The main goal of my thesis was to look for methods to improve the knowledge regarding the rooftop PV fleet and ultimately to introduce new methods for accurately estimating the PV power production.

In the thesis, we've introduced a method for constructing a large scale registry of rooftop PV systems and demonstrated that we could estimate the rooftop PV power production with relatively few information on these systems.

The last chapter of the thesis raises a lot of questions for improving current methods for PV power estimation, in particular since the task of fairly comparing the proposed method with the TSO current practices is very challenging. I also have in mind potential applications beyond France, for instance in coutries where the share of small scale PV is higher than in France.

List of publications

Publications in peer-reviewed journals

• Kasmi, G., Saint-Drenan, Y. M., Trebosc, D., Jolivet, R., Leloux, J., Sarr, B., & Dubus, L. (2023). A crowdsourced dataset of aerial images with annotated solar photovoltaic arrays and installation metadata. Scientific Data 10(1), 59.

Abstract: Photovoltaic (PV) energy generation plays a crucial role in the energy transition. Small-scale, rooftop PV installations are deployed at an unprecedented pace, and their safe integration into the grid requires up-to-date, high-quality information. Overhead imagery is increasingly being used to improve the knowledge of rooftop PV installations with machine learning models capable of automatically mapping these installations. However, these models cannot be reliably transferred from one region or imagery source to another without incurring a decrease in accuracy. To address this issue, known as distribution shift, and foster the development of PV array mapping pipelines, we propose a dataset containing aerial images, segmentation masks, and installation metadata (i.e., technical characteristics). We provide installation metadata for more than 28000 installations. We supply ground truth segmentation masks for 13000 installations, including 7000 with annotations for two different image providers. Finally, we provide installation metadata that matches the annotation for more than 8000 installations. Dataset applications include end-to-end PV registry construction, robust PV installations mapping, and analysis of crowdsourced datasets. The article can be accessed here and the Zenodo repository containing the dataset by clicking here: .

• Kasmi, G.; Touron, A.; Blanc, P.; Saint-Drenan, Y.-M.; Fortin, M.; Dubus, L. (2024) Remote-Sensing-Based Estimation of Rooftop Photovoltaic Power Production Using Physical Conversion Models and Weather Data. Energies 17(17), 4353.

Abstract: The global photovoltaic (PV) installed capacity, vital for the electric sector’s decarbonation, reached 1552.3 GWp in 2023. In France, the capacity stood at 19.9 GWp in April 2024. The growth of the PV installed capacity over a year was nearly 32% worldwide and 15.7% in France. However, integrating PV electricity into grids is hindered by poor knowledge of rooftop PV systems, constituting 20% of France’s installed capacity, and the lack of measurements of the production stemming from these systems. This problem of lack of measurements of the rooftop PV power production is referred to as the lack of observability. Using ground-truth measurements of individual PV systems, available at an unprecedented temporal and spatial scale, we show that by estimating the PV power production of an individual rooftop system by combining solar irradiance and temperature data, the characteristics of the PV system inferred from remote sensing methods and an irradiation-to-electric power conversion model provides accurate estimations of the PV power production. We report an average estimation error (measured with the pRMSE) of 10% relative to the system size. Our study shows that we can improve rooftop PV observability, and thus its integration into the electric grid, using little information on these systems, a simple model of the PV system, and weather data. More broadly, this study shows that limited information is sufficient to derive a reasonably good estimation of the PV power production of small-scale systems.. The article can be accessed here.

Workshops (peer reviewed)

• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2023). Assessment of the Reliablity of a Model's Decision by Generalizing Attribution to the Wavelet Domain. In XAI in Action: Past, Present, and Future Applications. Link.
• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2023). Can We Reliably Improve the Robustness to Image Acquisition of Remote Sensing of PV Systems?. In Tackling Climate Change with Machine Learning workshop at NeurIPS 2023. Link.
• Kasmi, G., Dubus, L., Blanc, P., & Saint-Drenan, Y. M. (2022). Towards unsupervised assessment with open-source data of the accuracy of deep learning-based distributed PV mapping. In MACLEAN: MAChine Learning for EArth ObservatioN Workshop co-located with the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD 2022). Link.

Oral presentations

• Kasmi, G., Touron, A., Blanc, P. Saint-Drenan, Y.-M., Fortin, M.,& Dubus, L. (2023). Enhancing regional PV power estimation using physics-based models, solar irradiance data and deep learning. In International Conference in Energy and Meteorology (ICEM), Padova, Italy . Slides.
• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2022). Leveraging earth observation data and deep learning to estimate the PV output in France. In MACLEAN Workshop @Cap/RFIAP, Vannes, France . Slides.
• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2022). Assessment of the potential of Earth observation data and deep convolutional neural networks to improve the estimation and forecast of the solar power production in France. PVPS Tasks 16 experts meeting, 2022, Sophia-Antipolis, France.

Preprints

• Trémenbert, Y., Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2023). PyPVRoof: a Python package for extracting the characteristics of rooftop PV installations using remote sensing data. arXiv preprint arXiv:2309.07143. Link.

Posters

• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2022). Assessment of the potential of Earth observation data and deep convolutional neural networks to improve the estimation and forecast of the solar power production in France. In 4th MADICS Symposium, Lyon, France.
• Kasmi, G., Dubus, L., Saint-Drenan, Y. M., & Blanc, P. (2021). Solar Array Detection on Aerial Photography Based on Convolutional Neural Networks: Image of the Solar Array Characteristics and Image Backgrounds on the Out-of-domain Generalization. In SophIA Summit, Sophia-Antipolis, France .

Peer review

Journals

• Solar RRL
• Scientific Data

Conferences

• Ethics reviewer for the Benchmark and Datasets Track at NeurIPS 2024

Workshops

• XAI in action workshop (NeurIPS 2023)
• Tackling Climate Change with Machine Learning workshop (ICLR 2024)
• Women in Computer Vision workshop (ECCV 2024)
• Interpretable AI: Past, Present and Future workshop (NeurIPS 2024)

Miscellaneous works

• Kasmi, G., Dubus, L, Saint-Drenan, Y.-M. & Blanc, P. Looking for a frequency-based principle to predict the sensitivity of convolutional neural networks to Gaussian image perturbations.

This work in process was presented during the PhD Forum at ECML-PKDD 2022. It it a snapshot of our early attempts to use Fourier theory to explain the (lack of) robustness of a CNN classifier. This work later led to the WCAM. The manuscript is accessible here and the slides of the presentation here.