| Popis: |
In many tropical countries, where family farming predominates, such as on Réunion, the large number, the small size of the fields, and the heterogeneity of the landscapes make it very difficult to map the land use. However, the availability of remote-sensing data, accessibility to parallel computing machines, and recent developments in artificial intelligence have progressively led to the popularity and success of deep-learning-based methods in remote sensing. Maps of current land cover and land use based on remote sensing, even if not as accurate as required for legal maps, offer a quick and effective solution for agricultural production analysis or prediction, land management, or food security warnings, especially in tropical countries where very few observations and agricultural data are available. A deep-learning approach is proposed to generate land cover and land use from a very high-resolution satellite image. With observations covering two crop years, we built a ground-truth database from 5000 field observations, based on a 3-level typology. We used it to train a multi-scale, multi-level convolutional model that is carefully optimized to maximize the accuracy. A hierarchical strategy is built in the prediction process to ensure the consistency of the most detailed typology. The inference is then achieved on a Pleiades image (spatial resolution of 0.5m/pixel) covering the territory of Réunion. The proposed approach shows better accuracy than traditional machine-learning and state-of-the-art methods based on multi-source satellite imagery. The results show the efficiency of artificial intelligence in exploiting the massive geospatial earth observation data provided by new generation satellite sensors for agriculture monitoring. In the future, innovative unsupervised or combined deep-learning algorithms may yield even better results and pave the way for regular production of land use maps at a very low cost, even in small-scale agriculture systems. |
