We used the Convolutional Neural Network (CNN) method to create a shrub cover map from aerial imagery of the Jornada Experimental Range (JER) and Chihuahuan Desert Rangeland Research Center (CDRRC). We used National Agriculture Imagery Program (NAIP) aerial imagery acquired on 21 May 2011 to generate the shrub cover map. NAIP images have a 1m spatial resolution and four spectral bands (red, green, blue, and near-infrared). We used Convolutional Neural Network (CNN) for image classification (into shrub and non-shrub pixels). Specifically, we used a U-Net architecture for our CNN model. The U-Net model was implemented using the individual spectral bands of the NAIP imagery and the Normalized Differential Vegetation Index (NDVI) as inputs. The U-Net model was trained using random sample patches from a shrub cover map developed independently earlier for the Jornada Basin Long-Term Ecological Research (LTER) (Ji et al. 2019). A total of 13,000 samples, each has a size of 128 by 128 pixels, were generated and were then split into training (~70%) and validation (~30%) sets. The training images were normalized before training, which helped to cope with varying illumination conditions in the imagery. The training data were also extended by generating additional artificial and transformed training images (horizontal and vertical flipping; piecewise affine, perspective transformation; and linear contrast enhancement). The training data were used for the gradient-based optimization of the U-Net parameters and the training progress was monitored by computing the performance on the validation data. Model training achieved a per pixel test accuracy of 89.6%.
The shrub cover map used for model training is described and published in this journal article:
Ji, Wenjie, Niall P. Hanan, Dawn M. Browning, H. Curtis Monger, Debra P. C. Peters, Brandon T. Bestelmeyer, Steve R. Archer, et al. “Constraints on Shrub Cover and Shrub–Shrub Competition in a U.S. Southwest Desert.” Ecosphere 10, no. 2 (2019): e02590. https://doi.org/10.1002/ecs2.2590.
And the training data maps are published in the previous version of this data package (knb-lter-jrn.100.1) at:
https://doi.org/10.6073/pasta/2bbee949ad08c7feb1d5cec6570b65b8
