Caractérisation et suivi à long terme de la phénologie des couverts agricoles par utilisation des modèles d’apprentissage automatique et les données spatiales à moyenne résolution spatiale

Abstract

Changing land-use patterns are of great importance in environmental studies and critical for land use management decision-making over farming systems in arid and semi-arid regions. Unfortunately, ground data scarcity or inadequacy in many regions can cause large uncertainties during the characterization of phenological changes in arid and semi-arid regions, which can hamper tailored decision-making towards the best agricultural management practices. Alternatively, state-of-the-art methods for phenological metrics extraction and long-time series analysis techniques of multispectral remote sensing imagery provide a viable solution. In this context, this thesis aims to assess the relevance of phenology data and machine learning algorithms to characterize the changes over farming systems. It also investigates the strength and validity of phenology data combined with climate data to provide the first pheno-climatic classification of Morocco. To this end, four farming systems (FS) (Fallow (FA), Rainfed area, (RA) Irrigated Annual Crop (IAC), and Irrigated Perennial Crop (IPC)) in arid areas of Morocco were studied based on 13 phenological metrics (PhM). These metrics were derived from large MODIS-NDVI time-series between 2000 and 2020. For classification and change analysis purposes, 3 machine learning algorithms and a pixel-based change analysis method were investigated. Besides, long time series of climatic data (i.e., rainfall data and land surface temperature) and phenological metrics were used to produce the first pheno-climatic classification of Morocco at a scale of 500 m. The classification overall accuracy over the Beni Mellal-Khenifra region reached 88%, with a kappa coefficient of 0.83 and values of F1-score greater than 0.76. However, by comparing, the accuracy of the three classifiers (i.e., Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbour (K-NN)), the RF method showed the best performance with an overall accuracy of 0.97 and kappa coefficient of 0.96. Variations in FS have been found to be linked well with other indicators of local agricultural land management, as well as the historical agricultural drought changes over the study area. Results showed a significant dynamism of the plant cover linked to the behaviour of farmers who tend to cultivate intensively and to invest in high-income crops. More specifically, a relevant variability in fallow and rainfed areas closely linked to the weather conditions was found. In addition, a significant lag trends of the start (-6 days), end (+3 days) of season was found, and which indicate that the length of the season was related to spatio-temporal variability of rainfall. This study has also highlighted the potential of Multitemporal moderate spatial resolution data to accurately monitor agriculture and better manage land resources. In the meantime, for operationally implementing the use of such work in the field, we believe that it is essential to consider the perceptions, opinions, and mutual benefits of farmers and stakeholders to improve strategies and synergies whilst ensuring food, welfare, and sustainability.