Crop Classification from Multi-Temporal and Multi-spectral Remote Sensing Images

The number of satellites, equipped with various sensors, aiming to observe agricultural activities have been progressively increasing. Satellite technology advances have enabled the acquisition of multispectral images of a region with small temporal intervals. Consequently, changes over a region can be observed, yield forecast can be made and the type of crops can be determined. In this work, it is aimed to classify 13 different crops by processing the multi temporal and multispectral data consisting of surface reflectance values. To this end, a siamese recurrent neural network structure, that processes time series information with deep metric learning approaches and providing easier classification, is proposed. A convolutional neural network that processes the multi temporal and multispectral information like an image is proposed to reduce the effect of class imbalance problem. These models are then combined under an ensemble neural network structure in order to leverage both networks' strengths. The proposed method outperforms other studies on the literature on BreizhCrops dataset.