Chili Crop Detection using Satellite Imagery and Machine Learning

Challenge Completed!

A global team of 50 AI engineers collaborated to estimate the total cultivated area and growth stages for fields growing chili for a given district in India states.

The partner for this project, Farm-Hand, is a software and data analytics company (UK and India based) that uses satellite data, and AI/ML alongside initial farmer-led data to obtain field-level insights.

The problem

The Indian Farming sector is dominated by the smallholder farmer who on average is managing an area of 2 acres. Many smallholder farmers are affiliated with aggregator companies, whose role may include access to debt finance, provision of seeds/solutions (e.g. fertilizer), farm management practices, and access to the market. The project partner, Farm Hand, is providing its farm management platform to a number of Aggregator Companies who work with chili farmers in the States of Karnataka and Andhra Pradesh in Southern India. Approximately 30 meta-data farm/field attributes are collected, as each farm is onboarded to the platform. Paramount among these with respect to this project are growing season dates for chili production over the last 12 months. In addition, field boundaries are defined and stored as KML files.

The data

Crop growth progression between the dates defined, for the fields identified can be tracked using remotely sensed data from Sentinel Satellite. The Copernicus Open Access Hub provides complete, free and open access to Sentinel-1, Sentinel-2, Sentinel-3, and Sentinel-5P user products, starting from the In-Orbit Commissioning Review (IOCR). Sentinel Data is available via the Copernicus Data and Information Access Services (DIAS) through several platforms. The Institute of Crop Science and Resource Conservation (INRES) at the University of Bonn maintains a Remote Sensed Index Database (IDB). The IDB provides a quick overview of which indices are usable for a specific sensor and a specific topic. In addition to these indices, ESA has developed a number of models, based on ground truth data for eliciting specific features of vegetative growth from remote sense data that include Leaf Area Index (LAI), Fractional Cover (FCOVER), and Fraction of Absorbed Photosynthetically Active Radiation(FAPAR). These indices have been used by many research groups for detecting field boundaries and for creating crop classification algorithms, for instance, Orynbaikyzy et al., 2019, Saini et al., 2018.

Farm Hand provided a dataset containing an initial 300 individual entries with the following information:

Defined field boundaries of a field where chili cultivation has taken place.
The sowing ad harvest dates of at least one chill crop cycle for each field.

Over the duration of the project, additional farms have been onboarded to the platform at an approximate rate of 100 farms per week.

The project outcomes

The objective of this project was to build a machine learning model to detect chili crops and the boundaries of farms containing chilies – for a specific region of India.

The initial task was to gather and understand research that had been conducted for similar projects. Ground truth data, that included farm boundaries and crop cycle information was provided by FarmHand. Satellite imagery was obtained from Sentinel Hub and Google Earth Engine. The model development consisted of two main tasks:

Field Boundary Detection/Delineation Modelling (FBDM), and
Crop Classification using Clustering

For FBDM, the team followed 2 different approaches based mainly on François Waldner’s authored ResUNet-a and FracTAL ResUNet architectures. Both approaches followed Sherry Wang’s paper on field delineation in smallholder farming systems with transfer learning.

For Crop Classification, we used the k-means clustering method where the input was four-time series data sets consisting of the following indices- NDVI, SAVI, NDWI, and EVI (see discussion below) These were calculated from Sentinel-2 images downloaded for the chili crop cultivation cycle (July through March). We used Google Earth Engine to download images (utilizing only images with less than 10% cloud cover) for the cultivation cycle during the year 2019-20. By overlaying the known farm boundaries from FarmHand data, we could improve our confidence in identifying and predicting other farms we believed to contain chili.