Identification d’aires mellifères productives par une méthode d’analyse d’images satellitaires et technique d’intelligence artificielle

The global decline in honeybee (Apis mellifera) populations, driven by climate change and habitat degradation, poses significant problems as these insects are crucial for producing honey, wax, and other products. Honeybees also play a vital role in pollination, supporting biodiversity, increasing agricultural yields, and contributing to the economy. For instance, in 2021, the value of honeybee pollination in Canada was estimated at over 3 billion CAD, highlighting their importance to the agricultural sector and the broader economy. Nonetheless, the survival of honeybees is increasingly jeopardized by factors such as climate change, habitat loss due to the expansion of farmland, and pesticide exposure. A promising strategy for proposing solutions to this in honeybee populations decline is to identify areas where these stress factors are mitigated, using beekeeping potential map. These maps allow for the identification the areas that are best for beekeeping, where the productivity can be maximized, different risks can be reduced. Such maps can assist beekeepers and government agencies in making informed decisions about where to establish apiaries, ultimately contributing to the protection and support of honeybee populations. Implementing these tools is challenging due to the need for integrating diverse factors that vary across spatial and temporal scales, necessitating mathematical modeling techniques. These models must seamlessly account for the complex interplay of these factors within a unified framework. In this study, we use fuzzy inference systems under two paradigms: expert knowledge and data mining modeling, to assess beekeeping potential. Initially, a Mamdani-type fuzzy inference system was developed, integrating expert knowledge with multi-source geospatial data in a hierarchical manner. This hierarchical model was applied in an area located in southern Quebec, demonstrating both reliability and effectiveness. Building upon this hierarchical fuzzy model, we proposed a rules simplification framework that permitted us to improve his interpretability. The results of the simplified model show a significant improvement in interpretability, reducing the number of rules from forty-three to six and variables from thirteen to five. This simplification process enhanced the model's predictive capability, improving the mean squared error by 30% compared to the original flat fuzzy system. In the following model, we introduce an innovative approach for predicting beekeeping potential areas by employing an adaptive neuro-fuzzy inference system with subtractive clustering. The study incorporates weather variables and land cover quality as key factors to assess the suitability of an area, using hive mass as a proxy. This approach demonstrated strong predictive capabilities, achieving a high correlation (R = 0.87) during the testing phase. Based on the analysis of these three models, it emerges that the land cover quality variable plays the most significant role in determining the beekeeping suitability of an area. In the final step of our research work, we utilized very high-spatial-resolution satellite images to estimate the floral diversity and abundance of beekeeping plants in fallow and grazing areas. A spectral unmixing algorithm was applied to multispectral very high-spatial-resolution images (Neo-Pleiade, WorldView-3, Planet SuperDove) to identify bee plants (verge d’or scientific name Solidago canadensis and eupatoire maculée scientific name Eutrochium maculatum). These plants were observed in the field near Deschambault, Quebec, around the time of the satellite’s pass. While not all sensors yielded conclusive results, the spectral unmixing technique, when applied with very high-spatial-resolution satellite images, demonstrated promising potential under suitable conditions.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.