Image for post
Image for post

Seeing Through the Clouds

Predicting Vegetation Indices Using Radar Imagery

Descartes Labs
Aug 26 · 7 min read

By Terry Conlon — Ph.D. Student in the Sustainable Engineering Lab at Columbia University. Terry completed the following work as a Descartes Labs intern this summer, where he worked with Rose Rustowicz to apply deep learning methods to timeseries of satellite imagery.

Despite numerous advances in the collection and distribution of satellite imagery — including those that allow sensing at sub-meter resolutions, rapid orbital periods, and public access to petabytes of data — one fundamental problem remains: how do you deal with clouds? In many parts of the world, persistent cloud cover prevents reliable information extraction from optical imagery, rendering many monitoring efforts dead on arrival.

Even though image contamination from clouds is both inevitable and ubiquitous, there is one silver lining: Cloud-cover does not impact all satellite imagery products! Certain data collected from radar sensors, such as those aboard the Sentinel-1 satellites, are not meaningfully affected by weather. As a result, scientists have been experimenting with methods of transforming non-cloudy radar images into optical ones, using advances in deep learning to extract hidden relationships between datasets: Recent papers have explored using either Sentinel-1 synthetic aperture radar (SAR) or interferometric synthetic aperture radar (InSAR) data layers to monitor vegetation phenology [Frison et al. (2018)], map crop types [Mestre-Quereda et al. (2020)], and generate optical imagery [Schmitt, Hughes, & Zhu (2018); Toriya, Dewan, & Kitahara (2019)].

At Descartes Labs, I built upon this work over the summer, using both SAR and InSAR timeseries to predict vegetation indices (VI) in a region surrounding Fresno, California, and across an area of the Great Plains that enclosed portions of Iowa, Nebraska, and South Dakota; the exact boundaries of these regions are shown in Figure 1.

Image for post
Image for post
Figure 1: Regions of interest for VI prediction using radar data, seen in the Descartes Labs Viewer.

SAR measures two-dimensional surface backscattering, whereby polarized radar waves are transmitted onto a target, and the return waves are collected and recorded. On the Sentinel-1 mission, satellites emit vertically polarized waves and receive waves of both horizontal (denoted vh) and vertical polarizations (denoted vv). Reflection off vegetated surfaces can cause vertically polarized radar waves to return horizontally polarized (vh), although this process is highly dependent on the type of vegetation and its growth stage.

By comparing SAR backscatter measurements taken on two separate satellite passes, scientists can create an InSAR interferogram. The coherence layer of an interferogram measures the consistency of the scattering processes in the two SAR collects. Surface cover land types that demonstrate little change over time, such as buildings or pavement, will display high coherence; land cover types that produce different amounts of backscatter due to changing surface geometry will display low coherence values. Vegetation falls in this second category: Due to weather conditions or growth, leaves and branches can move by many wavelengths between passes, resulting in varying levels of backscatter over time.

SAR and InSAR coherence layers, neither of which are appreciably affected by cloud cover, are shown in the first two columns of Figure 2. Column A presents vh backscatter in the red and green bands of the images and vv backscatter in the blue band; accordingly, pixels with high vh values look yellow and pixels over land cover types that do not change the polarization of the radar signal appear blue. Column B shows InSAR coherence: low coherence pixels are darker than their high coherence counterparts. In both displayed scene settings, we see alignment between the vh backscatter layers in Column A, the pixels with low coherence values in Column B, and the Sentinel-2 derived normalized difference vegetation index (NDVI) layer, which is presented in Column D. For reference, Column E presents the settings in true color RGB. Column C shows our NDVI predictions, generated with a modified Pix2Pix generative adversarial network (GAN), which is discussed in detail below.

Image for post
Image for post
Figure 2: Column A presents SAR backscatter, with the vh layer in both red and green bands and the vv layer in the blue band; Column B presents single band images containing InSAR coherence; Column C presents NDVI predictions from our model; Column D presents NDVI derived from Sentinel-2 measurements; Column E presents true-color RGB images of the landscapes. Row 1 presents images collected around June 15, 2019 for a setting outside of Fresno; Row 2 presents images collected around October 7, 2019, for another setting outside of Fresno.

Since their introduction in 2014, researchers have used GANs for a variety of different image generation tasks. In the case of paired image translation — a type of transformation that leverages the same image in multiple domains to extract cross-domain mappings — the Pix2Pix GAN architecture has successfully learned transformations in multiple use cases [Isola et al. 2017]. As our objective involved translating radar imagery into its optical counterpart, we modified a Pix2Pix to serve as our prediction model.

Our modified Pix2Pix used timeseries of SAR backscatter and InSAR coherence, along with Shuttle Radar Topography Mission slope measurements, as model inputs; it used Sentinel-2 derived VIs as targets. The input timeseries consisted of 4 timesteps (t’ = t .. t-3) 12 days apart, a duration which corresponds to the orbital period of a single Sentinel-1 satellite on a single ascending or descending pass. Similarly, coherence was calculated from Sentinel-1 passes 12 days apart. Input SAR timeseries were coregistered, spatiotemporally denoised, and converted to log-space; target VI layers (at time t’ = t) were clipped to the non-negative portion of their physical range. All input and target data were collected and organized with the Descartes Labs Platform.

We made an important modification to the baseline Pix2Pix training configuration in the form of an L1 crop loss: a loss proportional to the absolute difference between target values and predictions over cropland, determined by the USDA Cropland Data Layer. We imposed this loss to incentivize accurate predictions over cropland, as we felt that these were more important than predictions over non-cropped surfaces.

Predictions of enhanced vegetation index (EVI) using radar imagery, along with derived EVI from the Sentinel-2 images closest to them in time, are shown in the GIF below.

Image for post
Image for post
Figure 3: VI predictions and derived values for an area of interest (AOI) in Central Iowa; predictions range through 2019.

We can see in this GIF an overall alignment between our predictions and the VIs that would be expected at that time of year: Crops in Iowa cycle annually, with growth beginning after May and continuing through the summer until harvest around October. We also see a number of Sentinel-2 scenes where cloud cover renders the image meaningless. In these cases, our predictions provide information that is otherwise unavailable.

More generally, we have five primary findings from our work:

  1. EVI predictions were more accurate than NDVI ones. With our EVI models, we encountered lower average pixel errors and fewer instances where predictions seemed highly unrealistic or did not capture the scene structure. It is likely that NDVI saturation is partly responsible for the larger discrepancies between NDVI model predictions and target measurements; as EVI does not saturate as easily, different input layer combinations are not paired with similar target values as often, making the radar-to-optical mapping easier to extract.
  2. A crop-specific L1 pixel loss offered slight benefits. The L1 crop loss caused marginal improvements in prediction accuracy over cropland. This finding held when we tailored the loss to a specific land-cover type: Applying a L1 loss over only corn fields caused a slight improvement to VI predictions for corn.
  3. Not all predictions performed equally. Predictions over certain land cover types — like corn and soybeans — were more accurate than those over others, like alfalfa.
  4. Training and validating over flat landscape was key. Because SAR backscatter is highly sensitive to topography, only collecting imagery over flat landscapes resulted in better predictions.
  5. Recurrent layers did not improve model performance. Adding recurrent layers to the Pix2Pix architecture in order to explicitly capture the temporal relationship of input timeseries did not increase prediction accuracy.

With the Descartes Labs Platform, we were also able to save our predictions for future retrieval, accessing them as we would for any other satellite imagery product. Figure 4 shows an example of how to retrieve predicted VI layers aggregated over a hand-drawn geometry and inspect the resultant timeseries, all in a Jupyter notebook.

Image for post
Image for post
Figure 4: Interactive display of predicted VI layers in a Jupyter notebook. Predictions (blue) and Sentinel-2 ground-truth (orange) are shown for NDVI in the top half of the pop-up plot and for EVI in the bottom half. Predictions are made for the entirety of 2019. The plot loading process is abbreviated by 8 seconds to condense the GIF.

In this GIF, VI predictions are shown for the yellow AOI. After selecting a portion of this region using a drawing tool, we can see VI predictions and Sentinel-2 derived measurements compared side-by-side.

While the results shown here demonstrate a baseline ability to convert radar imagery into vegetation indices, other aspects of the translation process deserve further attention. Future work should investigate model transferability: Can one trained model be applied in numerous settings, or is location-specific training required? Furthermore, these translation models will provide greater value if they can be reliably deployed over topographically complex terrain. By focusing on these open questions, researchers have the opportunity to solve one of remote sensing’s fundamental problems: how to see through the clouds that occlude optical imagery.

Thanks Terry! Stay tuned for more featured work by our scientists, interns, and users.

Check out our platform page, view our webinar or drop us a line to discuss how the Descartes Labs Platform can help accelerate your work.


Explore posts from the Descartes Labs team

Medium is an open platform where 170 million readers come to find insightful and dynamic thinking. Here, expert and undiscovered voices alike dive into the heart of any topic and bring new ideas to the surface. Learn more

Follow the writers, publications, and topics that matter to you, and you’ll see them on your homepage and in your inbox. Explore

If you have a story to tell, knowledge to share, or a perspective to offer — welcome home. It’s easy and free to post your thinking on any topic. Write on Medium

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store