dPEN: Deep Progressively Expanded Network for Mapping Heterogeneous Agricultural Landscape using WorldView-3 Satellite Imagery


Accurately mapping heterogeneous agricultural landscape is an important prerequisite for agricultural field management (e.g., weed control), plant phenotyping and yield prediction, as well as ecological characterization. Compared to traditional mapping practices that require intensive field surveys, remote sensing technologies offer efficient and cost-effective means for crop type mapping from regional to global scales. However, mapping heterogeneous agricultural landscape is a challenge because of diverse and complex spectral profiles of crops. We propose a novel deep learning method, namely deep progressively expanded network (dPEN), for mapping nineteen different objects including crop types, weeds and crop residues, in a heterogeneous agricultural field using WorldView-3 (WV-3) imagery. To assess the mapping accuracy of dPEN, we created a calibrated WV-3 dataset with the corresponding ground truth. In addition, the suitability of visible/near-infrared (VNIR, 400-1040 nm) and short-wave infrared (SWIR, 1195 nm-2365 nm) bands of WV-3 to classification accuracy were examined and discussed in detail. To the best of our knowledge, this is the first effort to explore the significance of all SWIR bands in WV-3 for classification accuracy in a heterogeneous agricultural landscape. The results demonstrated that: (1) The proposed dPEN allows for building a deeper neural network from multispectral data which was the limitation of many convolutional neural networks; (2) dPEN was able to extract more discriminative features from VNIR and SWIR bands by producing the highest overall accuracy (OA: 86.06%) over competing methods such as support vector machine and random forest; (3) The inclusion of WV-3 SWIR bands greatly improved the classification accuracy; (4) SWIR bands were particularly beneficial to improve the classification accuracy of some individual classes such as weeds, crop residues, and corn and soybean during late developmental stages; (5) The red-edge band (705-745 nm) was identified as the most important band affecting the classification accuracy nearly 10%, whereas the coastal band (400-450 nm) provided the lowest contribution; and (6) SWIR-5 band (2145—2185 nm) contributed most to OA by enhancing it approximately 4% when combined with VNIR bands, while SWIR-1 (1195-1225 nm) yielded the lowest improvement (1.55%) for OA. These research outcomes provide useful information for efficiently mapping agricultural landscape, and indicate the potential practices of dPEN and contributions of spectral bands in WV-3 for plant phenotyping, weed control, and crop residue retention.


Civil, Architectural and Environmental Engineering


Funding for this work was provided by National Science Foundation ( IIA-1355406 and IIA-1430427), and NASA ( NNX15AK03H).

Keywords and Phrases

Convolutional neural network; Crops; Deep learning; Image classification; Residues; SWIR; VNIR; Weeds; WorldView-3

International Standard Serial Number (ISSN)


Document Type

Article - Journal

Document Version


File Type





© 2019 Elsevier Inc., All rights reserved.