NASA Harvest partners at Stanford’s Center on Food Security and the Environment (FSE) recently published a study on their efforts integrating lidar (Light Detection and Ranging) and optical earth observation (EO) data to improve crop type mapping in areas with low training data availability. While accurate crop type maps are essential for understanding the location and scale of production for the world’s major crops, the lack of data necessary for model training in many regions can complicate their creation. Compounding the problem is that models created using solely traditional data sources like optical EO data can have low transferability across geographic regions. The team from FSE sought to improve this transferability by introducing lidar data into their optical-based models. The team found that lidar profiles, when combined with optical imagery, can produce training labels for maize and enable crop type mapping at a 10 m resolution. The lidar profiles are able to distinguish maize from other crops with an 84% accuracy, and these profiles have a very high transfer accuracy of 82% between different regions, allowing for high resolution crop type mapping at a global scale.
Crop type mapping is an important technique in agricultural remote sensing. Knowing when major commodity crops like maize, wheat, rice, or soybeans are damaged in one area allows markets to prepare for potential disruptions. Likewise, knowing that a drought is harming a staple crop’s production in a food insecure region lets policymakers know that humanitarian measures may be necessary.
Remotely-sensed EO data assists in crop type mapping as it allows users to observe landscapes around the world at various points of the year and determine what is growing on the surface. To make these decisions, researchers use models that are reliant upon accurate labeling data. Unfortunately, this data can be lacking in many areas of the world. Further, models trained on data from one area may not have high accuracy when applied to a different spot on the globe.
NASA Harvest partners at FSE including Dr. David Lobell, Dr. Sherrie Wang, and Ms. Stefania Di Tommaso explored ways to overcome these shortcomings in the paper “Combining GEDI and Sentinel-2 for Wall-to-Wall Mapping of Tall and Short Crops”. The paper focuses on their use of EO data fusion.
EO data fusion refers to the combination of different kinds of EO data. This study blended optical data and lidar data. Optical data refers to the spectral reflectance, or the amount of light that is reflected off of the surface of an object, of the visible spectrum (e.g. red, blue, green). The optical data for this study comes from the European Space Agency’s Sentinel-2 platform. While commonly used in crop type models, optical data has some drawbacks. As mentioned above, optical data-based models made for one geographic area can have low transferability to a different area, or even the same area in a different year. Thus creating global-scale crop type maps, particularly at a high resolution, is difficult.
Lidar return pulse along GEDI instrument path (track) showing vertical distribution of vegetation. Source: UMD GEDI.
Lidar is another form of EO data, but instead of looking at spectral reflectance, it instead focuses on how far light travels before being reflected back. A pulse of light is sent from a lidar satellite platform, hits something on the surface of the planet, and reflects back to the lidar sensor. When the light hits a tall object like a skyscraper or a tree, the travel time is less than when it hits grass or a small building as the light is reflected back at a greater height. This allows researchers to create 3D renderings of the Earth's surface, called lidar profiles. Lidar data is therefore very useful in mapping things that have height or textures as notable characteristics, like an urban topography or a canyon’s layout. The team at FSE used data from the GEDI mission, a lidar sensor attached to the International Space Station with the original objective of observing and measuring the world’s forests. Given that the height difference between different crops tends to be much less than what is seen between various trees in a forest, the GEDI mission was not originally intended to be applied for agriculture. This study thus also serves as an opportunity to evaluate the potential for the GEDI mission to expand its scope to include agriculture.
The team at FSE used lidar to help differentiate between maize and non-maize crops. As maize is one of the tallest cultivated crops at more than 2 meters or 6.5 feet tall, it lends itself to be distinguished from shorter crops like soybeans or rice. Using optical-based cropland maps showing where cropland is and then layering a lidar profile to separate maize from non-maize crops allowed the team to create crop type maps showing where maize is being grown.
The data fusion technique was tested in geographically-diverse, major maize growing regions with a mixture of short and tall crops in the United States, France, and China. The team created an optical-lidar classification model for each study area, and the model was evaluated for various months within the growing season. The technique was able to correctly classify maize crops with high accuracy for each region that it was trained for. Each region had highest accuracies in different months (88% for September in China, 85% for July in France, and 91% for August in the United States), but all months showed overall accurate results. August was the best month overall as all models had accuracies over 83%.
Classification accuracies for locally-trained GEDI and Sentinel-2 models. Bars indicate mean GEDI accuracies for models trained in different months. Error bars show one standard deviation. The dashed line indicates accuracy of the S2 Local model in each region, which is 93% in China, 95% in France, and 95% in the U.S. Note that training sample locations for the Sentinel-2 model were the same as the training shot locations for GEDI for all three months.
Each trained classification model was then applied to the other two regions. None of the models were given new training data from their new region in order to test the accuracy of the model’s transferability. This analysis found that each of the models performed well in classifying maize crops in new regions without being given new data. While generally each model was strongest in the region it was originally trained for, each model had very similar accuracies in the other two regions it was tested in. In some cases, such as August performance in France, the model’s accuracies were indistinguishable.
Test accuracies for GEDI models when using different combinations of training and test regions and periods. Colors indicate the region where models were trained, with hatching indicating a model trained in the same region (but on different locations). Models trained in other regions typically perform similarly to those trained locally.
While these results show the potential of GEDI data for mapping maize, the GEDI mission is limited in scale as only a small portion of the planet is mapped. Because of this, the team at FSE explored how they could export the success of the GEDI labeling to create a wall-to-wall crop type map. They did this by training a Sentinel-2-based optical maize classification model with GEDI predictions. While optical-only classification models suffer accuracy decreases when transferred to different regions, training an optical model with the GEDI lidar data made the classification much more accurate outside original training locations. The team at FSE found that the optical-only model had an overall average accuracy of 64%, while the GEDI-trained optical model had accuracies for all three sites above 82%. As Sentinel-2 imagery has a spatial resolution of 10 m, these results show that GEDI lidar profile-trained models can create accurate, transferable, high resolution wall-to-wall crop type maps.
Ground truth crop type maps (a)–(c) compared with classification predictions of GEDI-S2 Transfer models (d)–(f). Prediction maps for China (d) and France (e) are created using models trained in the U.S. (84% accuracy for both). The prediction map for the U.S. was created using the model trained in France (86% accuracy).
These results show the promise of lidar in distinguishing between short and tall crops like maize and that this capability can be transferred across different regions of the world with high accuracy. As the authors note, if lidar data becomes more densely sampled across the earth’s surface than it currently is, it would provide a great supplement to maize mapping efforts. Its inclusion could dramatically decrease crop type mapping costs in terms of labor and money. Even with current levels of GEDI sampling however, the authors demonstrate that the mission’s lidar data allows for expanded mapping of maize in areas with little to no available training data. The demonstrated ability of GEDI data to be utilized for crop type mapping expands the original scope of the mission and is an incentive for the continuation and expansion of GEDI lidar collection.
Combining GEDI and Sentinel-2 for Wall-to-Wall Mapping of Tall and Short Crops is open access and can be read here