A Spatio-Temporal Deep Learning-Based Crop Classification Model for Satellite Imagery

Abstract:

Intelligent agriculture monitoring systems can improve farm productivity by relying on Earth Observation (EO) data to facilitate practical large-scale solutions with high precision. However, the scarcity of annotated EO data is a major challenge for such systems. In this work, a spatio-temporal Deep Neural Network (DNN) model is presented to classify crop types in small farms using time-series multispectral satellite imagery. Our model consists of a multi-layer Convolutional Neural Network (CNN) with a special pooling layer that takes into consideration small farm area, followed by a multi-layer Gated Recurrent Unit (GRU) network and a fully-connected layer for output classification. The proposed model processes input images on two stages. In the first stage, the CNN extracts spatial features from each image in the time-series data. In the second stage, the GRU network analyzes the spatial features of all images and extracts temporal features representing the input sequence. To overcome the issue of data scarcity, the model is trained using extensive data augmentation techniques. Despite being trained only on ~3k sequence of images, the proposed model achieved first place in the Radiant Earth’s Computer Vision for Crop Detection from Satellite Imagery Workshop Competition in ICLR2020 surpassing all models that depend on hand-engineered features and traditional Machine Learning (ML) classifiers.