PTI
New Delhi
Researchers have developed a new dataset aimed at helping AI systems acquire capabilities to monitor crops across vast agricultural areas, identify what is growing in individual fields and even estimate when crops are sown, harvested and how much they produce.
A multi-satellite dataset has been developed by researchers from the Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), along with the MS Swaminathan Research Foundation, that combines satellite imagery with field-level information collected from agricultural plots, in a study called ‘SICKLE: A multi-sensor satellite imagery dataset annotated with multiple key cropping parameters’.
The dataset, called SICKLE (satellite imagery for cropping annotated with key parameter labels), is designed to help machine-learning models understand multiple aspects of crop cultivation rather than simply identify vegetation in satellite images.
The researchers said such datasets are crucial for developing AI-based agricultural technologies because while satellite imagery can cover large areas, the availability of detailed, field-level information needed to train and test AI models remains limited.
By bringing together imagery from three satellites, namely Landsat-8, Sentinel-1 and Sentinel-2, along with information on crops and farming practices, the dataset allows researchers to study crop identification, crop growth stages and yield prediction using the same agricultural plots.
The dataset contains around 2,370 season-wise samples from 388 agricultural plots and about 209,000 satellite images. It includes information at resolutions of 3 metres, 10 metres and 30 metres, allowing researchers to examine agricultural fields at different levels of detail.
For AI systems, this combination could be particularly significant, researchers said.
Instead of relying only on what a satellite image looks like, a model can be trained against information about what was actually cultivated in a field and when key stages of cultivation occurred, they said.
The researchers, which included Depanshu Sani, Sourabh Saini, Harsh Kumar Agarwal, Saket Anand and Gaurav Arora from IIIT Delhi, have used the dataset to benchmark three tasks: identifying crop types, estimating crop phenology and predicting yields.
Crop phenology refers to important stages in a crop’s life cycle, including sowing, transplanting and harvesting. In the dataset, these events are linked to individual plots, providing AI models with information that can help them learn how crops change over time.
The researchers also tested whether information from different satellites could be combined to improve predictions. The three satellites capture different types of information, including optical, thermal and microwave observations, and revisit agricultural areas at different intervals.
Such multi-source data could potentially allow agricultural monitoring systems to continue drawing information even when one type of satellite observation is limited. For instance, optical imagery can be affected by cloud cover, while radar observations are less sensitive to clouds.
The research points towards a model in which farmers, researchers and policymakers could eventually use satellite-based AI tools to obtain crop-related information over much larger areas than can be covered through conventional field surveys.
For policymakers, such systems could support large-scale monitoring of cropping patterns and agricultural productivity, while researchers could use them to study how crops respond to changing conditions. The dataset could also help in developing technologies for crop mapping and yield estimation at the field level.
However, the researchers noted several challenges, including the limited availability of field surveys, uncertainty in farmer-reported information and the difficulty of accurately identifying small agricultural plots in lower-resolution satellite imagery.
The SICKLE dataset was developed using agricultural plots in Tamil Nadu’s Cauvery Delta, a major rice-growing region. Field information collected for the study included crop type and variety, growing season, yield and key crop dates.
The dataset covers the period from January 2018 to March 2021 and includes 21 crop types across four districts. Of the samples, 351 paddy samples from 145 plots contain detailed information on paddy variety, growing season, phenology and productivity.
