Parcel-Scale Crop Type Classification Using Tile-Slice-Based Spatial-Temporal Convolutional Networks

Objectives Parcel-based crop classification utilizing multi-temporal satellite images is essential for precision agriculture. However, it is a significant challenge to explore the multi-scale spatial information for identification of crop types from remote sensing images.

Methods This paper proposes a tile-slice-based spatial-temporal (TSST) method for parcel-scale crop type classification using multi-temporal Sentinel-2 images. The core to the proposed method is the combined use of tile-slice-based feature representation of parcels and a deep spatial-temporal convolutional network. First, according to crop types and parcels, spatial tiles were sliced to produce a tile-slice-based training sample dataset and a tile-slice-based prediction dataset. Then, the deep spatial-temporal convolutional network was established to estimate crop-type probabilities for tile-slice-based prediction dataset. Finally, tile-slice-based probabilities were fused within parcel polygons to generate the final crop type maps.

Results In the study area in France, the proposed method is further discussed and validated through parcel-based time-series crop classifications using multi-temporal Sentinel-2 images. The overall accuracy, precision and F1 score are improved by 0.03, 0.02, 0.02, respectively. The classification results demonstrated great improvements over the comparison methods.

Conclusions We concluded from experiments and discussions that the proposed TSST method is effective for parcel-based crop classification using Sentinel-2 image sequences. The spatial division and slice in parcel-scale representation are benefit more for crop type identification of larger and spatially concentrated parcels. The proposed TSST method is lightweight but powerful for crop classification. These studies and findings will provide new ideas for parcel-based remote sensing time series analysis.