Mapping Rice Fields using Imagery Time Series
Updated: Sep 8, 2021
The Mekong River Delta is a very significant rice producing region of the world. Known as 'Rice Bowl', this region produces more than half of Vietnam's rice crop (Vietnam being the world's 5th largest rice producer and the 3rd largest rice exporter).
Mapping agricultural produce, especially a staple food such as rice, is very important. Not only does it employ a significant number of farmers (over 15 million rice farmers in Vietnam itself), but also the information on its growing patterns, productivity and health impacts planners, decision makers and marketers.
Just this year, due to limited internal supplies, Vietnam was forced to import rice from its trading rival - our country India.
Majority of my previous work on aerial imagery involved analyzing a single satellite image and drawing conclusions out of it. Occasionally, I compare the 'output' of analysis of two images to demonstrate the difference for better understanding.
For this application, we will use 15 radar images from the April to December 2020 to detect and understand growing behavior of Rice (Paddy) in the Mekong River Valley Delta of Vietnam.
(Much thanks to RUS Copernicus for supplying the tutorial and resources to perform this analysis)
Why do we need multiple images for our analysis? That too for rice fields, you may wonder. Consider the Optical imagery below -
Bing Virtual Earth Optical Imagery over Southern Vietnam (Mekong River Valley Delta)
Will you be able to identify whether the dark green areas are rice fields and the brown or light green areas are not rice fields at all? Upon giving due thought, most of us would likely respond with a negative (unless you're a farmer practicing in this region).
It is difficult to deduce even if I were to give you an added piece of information - 'the image was taken during the rice harvesting season'. After all, how will you be able to tell whether a particular field has been harvested for its rice or is pending to be harvested for its rice or is barren or has been harvested for some other crop whose harvesting season coincides with that of rice.
That being said, the accuracy of your deduction may improve if I were to give you multiple images from the harvesting season, however, it still wouldn't be reasonably satisfactory / accurate.
For this application, analyzing multiple satellite radar images helps to address the problem statement. While Radar imagery has several intrinsic benefits over optical images, from this use case point of view, its main benefit lies in how the radar waves (SAR - Synthetic Aperture Radar) interact with the structural properties of the land cover, especially that of rice fields / paddy.
15 Sentinel-1 SAR images from April 2020 to December 2020 chosen for this Rice Mapping exercise
You may be aware that rice fields are inundated i.e. under water during the planting / initial stage. Only during the growth stage does the rice plant emerge from the water and becomes visible.
As you will see later, it is this very structural property which helps us to distinguish a rice field from other agricultural fields or other types of land cover in general.
(Paddy image by PhotographyCourse on Unsplash)
When a SAR radar wave hits a flooded area the reflection (backscatter) bounces away whereas when it hits vegetation, the reflection bounces back / towards the satellite.
Backscatter Extract. Source: EO College
In the single, raw SAR image from 9th December, 2020 above - the darker regions represent areas with low backscatter i.e. these are likely to be inundated whereas the brighter regions represent areas with high backscatter i.e. these are likely to be vegetated.
Several processing steps are needed before the raw image is rendered to be useful for our agricultural application. Without delving into the technical aspects, you can visibly conclude that the post-processed image above has an enhanced B/W color diversity, has been subset-ted to our study area and is accurately geo-positioned when compared to the raw image.
Next, we create a temporal backscatter profile i.e. load all the 15 images together and see how a particular pixel (the coordinates of which are mentioned on the bottom left of the image above) has evolved over time.
From the graph on the left, you can observe that there are three distinctive troughs i.e. pixel values with low backscatter - in the 2nd image, in the 8th image and in the 13th image respectively. These correspond to images captured on 13th April, 18th July and 15th November 2020 respectively.
You can also observe that the troughs are usually followed with periods of increasing backscatter. In Vietnam, rice is generally cultivated thrice in a year - in spring, autumn and winter. Hence, we are in a position to take a reasonable guess - that the geo-pixel under review is used to cultivate rice i.e. is a paddy field.
We now proceed to enhance and emphasize this pixel property in a visually meaningful manner.
Which area is likely to be a rice field? It is fair to deduce that the areas which would have the highest difference in pixel backscatter values over the period of nine months are likely to be rice fields.
This is exactly what we proceed to do. From the 15 images, we will now compute a) the maximum backscatter value of the pixels, b) compute the minimum backscatter value of the pixels and compute the c) difference (max - min) backscatter value of the pixels. These 3 bands of information are used to create a single false-color composition / visual raster image which is displayed below -
Final Output: Rice Field Mapping Exercise
From a scale of dark violet to bright blue - the dark violet parts of the map represent areas with highest backscatter difference whereas the bright blues represent areas with the lowest backscatter difference over the course of the 15 images. Areas with high backscatter difference i.e. violet are likely to be rice / paddy fields.
Isn't this interesting?
[Please note, this is just a visual representation of the land cover and not a classification. We don't really know whether these are indeed rice fields or not but are making a reasonable guess basis the particular structural property of a rice field vis-à-vis those of other land cover types. To validate our visualization and turn it into a classification - we can use optical imagery and in-situ, supervised data to complement / enhance the findings obtained from analyzing radar imagery.]
Intelloc Mapping Services | Mapmyops.com is engaged in selling products which capture geo-data (Drones), process geo-data (Geographic Information System) as well as services (PoI Datasets & Satellite Imagery). Together, these help organizations to benefit from Geo-Intelligence for purposes such as operations improvement, project management and digital enabled growth. Write to us on email@example.com. Download our one-page profile here. Request a demo. Regards, Arpit