Deep learning for recognizing fine detailed crop types using time series satellite radar images

Abstract

<p>Crop identification and mapping using satellite remote sensing techniques is critical for agricultural monitoring and management. Distinguishing crops from satellite sensor image can be challenging given the irregular shape of fields, the complex mixture within smallholder farms, the variety of crops, and the frequent land use changes. The advances in satellite sensor techniques and classification algorithms allow us to acquire timely information on crop types at fine spatial scales. State-of-the-art research of crop classification involves the joint use of both optical and microwave satellite imagery.</p><p>Our current research aims to develop a recurrent neural network (RNN) for crop classification using Sentinel-1A time series backscatter images. The objectives of our study are to discriminate a wide variety of crops at fine spatial details and to increase the classification accuracy using time series images. A pilot study was performed on an area of the North-western Germany, for which we obtained the Land use registry across the growing seasons in 2018 as the ground reference data.  The area has a maritime influenced climate which is featured by warm summers and mild cloudy winters and flat terrain. The major crops identified include barley, rapeseed, rye, wheat, potatoes. We expect to observe five stages, which are planting, vegetative, reproductive, mature, and harvested stages, in the time-series pixel values of the crop types. The satellite images have been batch processed based on the ESA recommended procedures.  An initial time series analysis was performed on individual pixel values to detect and characterize the changes in different crop types. The next step was to explore the spatial distribution of the crops, i.e., the shape of the parcels. Image segmentation approaches were considered for dividing the image into small parcels for object-based image analysis rather than pixel-based classification. Because of the imbalance number of parcels, we resample pixel within parcels to avoid the problem of underfitting or overfitting.</p><p>Our modelling approaches are developed based on the Long Short-Term Memory (LSTM) deep learning models, which transform the temporal and dual-polarization input features into sequential hidden states, generate the output with scores, and then predict the crop types. This study can be extended to lands under similar climate and terrain conditions, and, with contribution to the understanding of the global agricultural system.</p>