Precision Agriculture     


 Kamal C Das photo photoShantanu Godbole photo photoJagabondhu Hazra photo photo Calvin (Cal) Swart photoLloyd A. Treinish photo

Precision Agriculture - overview


Last few years, IBM Research India is working on a special initiative on digital farming. One of the key objectives of digital farming/precision agriculture is to increase the farm productivity by increasing the visibility of agronomic states (such as soil moisture, crop health, weather, etc.) of farms, leveraging digitization, mobile, IoT and cognitive technologies. The team is working on developing a suite of solutions (pest risk prediction, plant disease/pest detection, crop identification, yield prediction, precision irrigation advisory services, automated farm boundary identification, cohort analytics, etc) leveraging the power of IBM Research’s big data platform called PAIRS. The team is also involved in leveraging the power of mobile smartphone technology to capture field images and applying deep learning and advance image analytics to bring in actionable insights on time.
In collaboration with IBM Watson Weaher and Media (aks TWC), the team has released multiple agri sercices -  "HDNDVI", "High Definition Soil Moisture", and "Crop Identification and Acreage Estimation", "Automated Field Boundary Identification" and "Farm Level Yiled Estimation". Current team os working on launching multiple value added services like short and long term commodity price prediction, Cohort Analytics, Synthetic Aperture Radar based crop phenology estimation, etc. 
The HD-NDVI service has the unique ability of blending different satellite data to derive a 30m-daily NDVI estimate. The objective of this service is to provide high temporal and spatial resolution NDVI for the region of interest over a period of interest. The service provides NDVI at a spatial resolution of 30m and temporal resolution of 1-2 days. This is achieved by blending satellite data from different satellite sources in a manner that allows estimating NDVI for required spatial and temporal resolutions. The service can be queried in four different ways depending on the region of interest and period of interest. The region of interest can either be a point in space or a polygon and period of interest can be the current day (nowcast) or an interval in the past (time series for a point or time lapse of images for polygon). The spatial and temporal extents of a query are the district and from January 1st, 2016 to current date respectively.  
Soil moisture has a strong influence on several precision agricultural applications such as identifying crop health, determining optimal irrigation schedule etc. The existing soil moisture products from remote sensing satellites as well as model simulations are either not available or difficult to obtain at a farm scale, rendering these products of little use in precision agriculture. Keeping this in focus, we developed a HD-Soil Moisture service to provide soil moisture data at a high resolution at any location of interest in near real time. This is achieved by blending remote sense satelyte data with data from physical land surface model which takes input as land type, vegetation type, atmospheric parameters, solar radiation and simulate the evolution of soil moisture for required spatiotemporal resolutions. 
The team has also developed a rich set of machine learning models to predict different pests and diseases for various crops such as paddy, potato, cotton, chilly, onion and tomato. We use different types of satellite data to monitor the crop health along with the highly accurate weather data to nowcast and forecast the chances of a particular disease to occur in an agricultural farm. Our models have the flexibility to work in different temporal and spacial resolution depending on the use case. We use novel parameter aggregation techniques to tune the models to work for different geographic regions. 
The Plant Pathology app uses cognitive machine vision capabilities designed specifically to identify plant diseases and deficiencies manifested on leaf surfaces by diagnosis of associated visual symptoms. Our approaches also leverage technical pathology information to provide strong diagnosis based on actual plant pathology.  
One of the major problems in agriculture is to perform the root cause analysis to identify the reasons behind the crop yield variations.  Many times,  even though farmers follow similar farming practices but there is a significant variation in the crop yield (30%-50%).   To address this, we built a cohort analysis solution that analyses multiple electronic field records of the fields along with environmental, weather, and plant biological conditions, irrigation management, etc., and compares different fields to identify similar subsets of fields.  Furthermore, our solution provides customized farming practices to improve crop yield by analyzing a set of attributes related to a cohort which helps in explaining the gap in the farming practices. 
A key factor underpinning the success of digital and precision agriculture is the availability of electronic field records (EFRs), which include precise boundaries of individual farm-holdings. EFRs can facilitate regular collection of field status information such as soil moisture, crops grown, pest and disease attacks, fertilizer application, and other field-level activities over crop growth cycles. Such monitoring can enable farmers to take better informed decisions and intervene more appropriately and thereby positively impact yield. The availability of accurate field details is also important for agencies responsible for drafting agricultural policies, financial organizations funding farm operations, and efficient administration of farm insurance. Given the criticality and urgency of digitization of field-level details, we have developed a solution for delineating the boundaries of individual farms from imagery covering croplands.  The solution uses a combination of image processing and machine learning techniques, is quite generic, and can work with a variety of aerial and satellite imagery, with resolution that can be as coarse as 10 m.
Another key task in the digital agriculture portfolio is automated crop mapping and acreage estimation, as early in the crop seasons, as possible. Such information can help  government organizations drafting agricultural policies and private enterprises dependent on agricultural produce formulate better plans and operations. This exercise has traditionally been manual. With the increasing availability of remote-sensed imagery, significant effort is underway to automate. We have developed machine learning models that leverage optical and radar based remote sensing products along with weather and other meta-information for mapping significant crops and estimating their acreage.
Contact: Shantanu Godbole <>, Jagabondhu Hazra <>


APIs available shortly