Leveraging Global Crop-Land Datasets to Improve Model Performance for Crop Classification in Data-Sparse Regions

Abstract:

Machine learning combined with remote sensing is a powerful tool, widely used to better understand where (crop masks) and what crops (crop-type masks) are being grown. Such data are critical for a large range of agricultural applications and for informing various food security and policy decisions. However, most methods are limited to spatially homogeneous areas, large scale agriculture and/or data-rich areas. Crop masks are particularly lacking in regions dominated by smallholder farming, where there is high spatial heterogeneity and collecting ground truth data can be challenging both financially and logistically.

In data-sparse regions, models can be developed and improved using data outside the region of interest, e.g., globally distributed datasets such as the crowdsourced GeoWiki dataset. We present a multi-headed classifier that leverages global datasets such as GeoWiki in addition to local labels to produce accurate crop and crop-type masks in data-sparse regions. Using time-series of 10-meter Sentinel-2 multi-spectral observations, we apply a base neural network to encode the input data (e.g., an LSTM or a temporal convolutional network). This encoding is then passed to two classification heads: a "global classifier", which learns from the GeoWiki dataset, and a "local classifier", which learns to classify pixels in the data-sparse region. This allows the model to learn from the GeoWiki dataset, while allowing it to focus on the data-sparse task. In addition, this enables global datasets to supplement data-sparse tasks even if the tasks are different (e.g., using the GeoWiki dataset, which has binary labels of crop/non-crop, to supplement a multi-class crop-type mapping task).

We validated this method with two data-sparse tasks: crop-type mapping in Kenya and crop/non-crop mapping in Togo. In both cases, the addition of the GeoWiki dataset leads to an improved performance, despite a large range in the labelled-dataset size (~1000 examples to ~10,000 examples).