Estimating the area of croplands and estimating each crop's area using remote sensing and machine learning

In this thesis, the area of croplands and each crop's area are estimated from a remote sensing hyperspectral image of the sentinel-2 satellite using one of the well-known machine learning algorithms, Maximum Likelihood Classifier (MLC). Wheat, beans, nigella, other plants, urban, and fallow are the classes we selected to classify the study area. Our study area locates in North-west Syria. We collected ground-truth GPS points, and we drew seventy polygons depending on them. Each polygon's pixels represent the same class. These polygons are used as input to MLC. We generated one hundred and fifty random GPS test points above the study area to calculate accuracy. The overall accuracy is 82% and (user accuracies, producer Accuracies) for classes are: wheat (83.3%, 90.9%), beans (90.9%, 71.4%), nigella (81.8%, 94.7%), other plants (79.3%, 65.7%), urban (82.4%, 73.7%), and fallow (80.9%, 92.7%). The estimated areas for classes are wheat 14.881 km2, beans 3.614 km2, nigella 13.34 km2, other plants 20.77 km2, urban 14.54 km2, and fallow 35.23 km2. The total uncultivated area is 49.77 km2, whereas the cultivated area is 52.61 km2. In this study, the MLC results are also compared with Support Vector Machine (SVM) and Random Trees (RT) results, and we found that the MLC algorithm gives much better results than other methods.