Post-season and in-season crop type classification for smallholder farms: reducing reliance on labeled data by learning latent features in unlabeled data

Abstract:

Cropland and crop type maps are critical inputs for agricultural analyses and a large range of decision and policy making. However, labeled examples for training machine learning models to predict these variables are scarce and can be difficult to acquire for regions dominated by smallholder farming and at risk of food security. This label scarcity is particularly problematic when training neural networks, which require a large number of labeled examples. To reduce the number of labeled examples required, we propose a multi-task classification approach that trains a classifier using a large dataset of unlabeled examples and a smaller dataset of labeled examples simultaneously. The architecture consists of a 1D-convolutional autoencoder and a shallow neural network classifier. The autoencoder transforms multi-spectral time series satellite observations into a latent feature representation and attempts to reconstruct the original input from this representation. The classifier takes this latent, or “encoded,” representation as input and makes either a binary prediction of crop/non-crop or a multi-class prediction of crop type. These two networks are trained end-to-end: the classifier is trained with cross-entropy loss while the autoencoder is trained with the combined classifier loss and reconstruction error. This approach also enables in-season classification since the autoencoder can be used to reconstruct or “fill in” incomplete time-series inputs. Using this method for cropland classification using Sentinel-2 multispectral observations in western Kenya, we show that using this approach improves classification accuracy compared to a classifier without autoencoder feature learning for small labeled datasets (<1000 labels). We have also demonstrated promising results using this method for in-season cropland classification in western Kenya. Our next steps are extending this approach for cropland and crop type classification in the rest of Kenya and southern Mali.