[Paper Review] Rapid Response Crop Maps in Data Sparse Regions

Hannah Kerner, Gabriel Tseng+6 more|arXiv (Cornell University)|Jun 23, 2020

Remote Sensing in Agriculture30 references17 citations

TL;DR

This paper presents a rapid, data-efficient method for generating high-resolution (10 m) cropland maps in data-sparse regions using a multi-headed Long Short-Term Memory (LSTM) network that combines global crowdsourced labels with limited local ground truth. The approach delivered a Togo cropland map in under 10 days during the COVID-19 pandemic, outperforming existing maps and enabling real-time humanitarian decision-making.

ABSTRACT

Spatial information on cropland distribution, often called cropland or crop maps, are critical inputs for a wide range of agriculture and food security analyses and decisions. However, high-resolution cropland maps are not readily available for most countries, especially in regions dominated by smallholder farming (e.g., sub-Saharan Africa). These maps are especially critical in times of crisis when decision makers need to rapidly design and enact agriculture-related policies and mitigation strategies, including providing humanitarian assistance, dispersing targeted aid, or boosting productivity for farmers. A major challenge for developing crop maps is that many regions do not have readily accessible ground truth data on croplands necessary for training and validating predictive models, and field campaigns are not feasible for collecting labels for rapid response. We present a method for rapid mapping of croplands in regions where little to no ground data is available. We present results for this method in Togo, where we delivered a high-resolution (10 m) cropland map in under 10 days to facilitate rapid response to the COVID-19 pandemic by the Togolese government. This demonstrated a successful transition of machine learning applications research to operational rapid response in a real humanitarian crisis. All maps, data, and code are publicly available to enable future research and operational systems in data-sparse regions.

Motivation & Objective

• To address the critical lack of high-resolution, up-to-date cropland maps in data-sparse regions, especially in sub-Saharan Africa, where smallholder farming dominates.
• To overcome the challenge of limited ground truth data for training machine learning models in regions with no accessible field data or infeasible field campaigns.
• To develop a machine learning method that enables rapid, operational cropland mapping for humanitarian crises, such as the COVID-19 pandemic.
• To demonstrate the transition of research-grade machine learning to operational use in real-world crisis response.
• To provide publicly available maps, code, and data to support future research and operational systems in data-sparse regions.

Proposed method

• A multi-headed LSTM network is trained to classify cropland presence in each pixel using a monthly time-series of Sentinel-2 multispectral satellite imagery.
• The model leverages two data streams: global crowdsourced labels from Geo-Wiki (tens of thousands of examples) and a small set of curated, locally labeled examples from photointerpretation of high-resolution satellite imagery in the target region.
• The global stream learns general features for identifying cropland across diverse geographies, while the local stream learns region-specific features for improved accuracy in the target area.
• The method uses a hybrid training strategy that combines global and local data to reduce reliance on extensive local labeling.
• Cloud-based computation via AWS enabled scalable and distributed processing, accelerating model training and inference across the entire country.
• The final model was deployed for country-wide inference in under 10 days, with file transfer and data access being the primary time bottleneck.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on a combination of global crowdsourced labels and minimal local ground truth produce high-resolution cropland maps in data-sparse regions?
• RQ2 How quickly can such a model be deployed to support rapid humanitarian decision-making during crises?
• RQ3 Can a relatively simple, shallow LSTM architecture outperform complex models in cropland classification with limited labeled data?
• RQ4 To what extent can existing public datasets like Geo-Wiki and Copernicus be used to benchmark and validate new cropland mapping methods?
• RQ5 What operational software engineering practices enable the transition from research to real-time crisis response?

Key findings

• A 10 m resolution cropland map for Togo was successfully generated and delivered to the government in fewer than 10 days, enabling rapid policy response during the COVID-19 pandemic.
• The model outperformed existing publicly available cropland maps (GFSAD and Copernicus) in accuracy and spatial resolution for Togo.
• The method achieved strong performance using only a small number of locally curated labels, significantly reducing the need for costly and time-consuming field campaigns.
• The use of a shallow LSTM architecture proved effective and efficient, reducing overfitting and training time compared to deeper models.
• Cloud computing (AWS) enabled scalable and distributed processing, allowing rapid inference across the entire country.
• The study demonstrated a successful transition from research to operational humanitarian response, with all data, code, and maps made publicly available to support future work.