Remote sensing is presently a critical component of agriculture technologies used increasingly by agribusinesses, governments, and other non-governmental organisations alike to map and monitor land use at scale. Remote-sensing data enables the tracking and optimisation of agricultural activities by various actors in the agroecosystem and is an essential input for data-driven smart agriculture. When combined with ground truth and other sources of information, remote-sensing data provides a comprehensive analysis of crop production activities.
Crop Detection Using CropIn’s Machine Learning Algorithms
Any in-depth analysis of crop production using agriculture technologies begins with the detection of crops using remote-sensing images derived from earth observation satellites. These satellites are positioned several hundred kilometers away from the surface and equipped with multispectral sensors to perform Earth imaging to capture high-resolution ground images in the visible near-infrared (VNIR) and shortwave infrared (SWIR) spectral zones. Some earth observation satellites have upto 13 spectral channels that help analyse the biophysical features of plants using vegetation indices, which are calculated as differences between two or more bands in the visible light (VIS), near infrared (NIR), and SWIR wavelengths.
From among at least a hundred different spectral indices, the NDVI (Normalized Difference Vegetation Index) is a vegetation index that is preferred the most by scientists to determine the condition, growth stages, biomass, and the yield estimate of crops. The index quantifies the presence of chlorophyll on land surface and helps assess if the observed region contains live green vegetation by evaluating the distinct colours or wavelengths of VIS and NIR sunlight that the plants reflect.
When mapping land-use, CropIn utilises the NDVI time series from Sentinel-1 (RADAR) and Sentinel-2 (optical) satellite-imagery data to distinguish croplands and crop types from other forms of land cover. The pre-existing knowledge of phenological cycles of the different crops facilitates their identification, which is then validated by CropIn’s existing rich pool of crop datasets. Historical information of the plot, also derived from satellite imagery, enables us to get a deeper insight into what the farmer has cultivated in the past, and this intelligence further lends to the data validation when we deploy the crop-detection model for a new plot of farm land. CropIn has developed a state-of-the-art, real-time crop identification system using a bag of deep learning models. The system uses 3D CNN and LSTM architecture to build individual models. While building the models, both pixel- and image-based approach are considered to make a more generalised system. While Sentinel 2’s optical data provides a wealth of information that enables crop detection on a bright, sunny day, if the satellite image is obscured by the presence of clouds, particularly during the rainy season (Kharif season), the system automatically switches to the models that adopt RADAR data (Sentinel-1).
To improve the accuracy of the crop detection model and to validate the result, we train them repeatedly by deploying them on both small territories and across a significantly larger expanse of area, such as a pin code or district. To make the performance uniform across different geographical locations, transfer learning techniques are used to build more region-specific individual models. For the larger area, the crop that the deep-learning engine detects is cross-verified with government data, if available, or with data collected using CropIn’s SmartFarm® for a particular season or crop. Another benefit of using the crop detection model, along with land boundary detection, is that it also helps identify the difference between the farmers assessment of their land area and the corresponding yield and what the algorithm detects. The novelty of CropIn’s system is that it can predict crops at any point in time, from planting to harvest, and there is no need to wait for full time series information.
CropIn has deployed the system across the whole of the Indian state of Maharashtra for predicting the crops with Sentinel-1 and Sentinel-2 data in the years 2018, 2019, and 2020. The overall performance for the deep-learning models based on government statistics and ground validation is between 60% and 80% depending on the regions, seasons, and the years when the crops were detected.
Figure: CropIn utilises the NDVI time series from Sentinel-1 (RADAR) and Sentinel-2 (optical) satellite-imagery data to distinguish croplands and crop types from other forms of land cover.
Precision Farming — The Gift of Agriculture Technology
Efficient farming systems that are guided by scientific and accurate data are made possible with several advancements in agriculture technologies. Crop detection capabilities, powered by remote sensing, facilitates producers and enablers of agriculture to optimise crop production with minimal human interference.
Farming and seed companies: Crop identification based on geo-tagged farm plots and defined land boundaries enables producers to estimate the yield more precisely and in real-time. It also helps producers to recognise signs of poor crop health caused by a disease or pests and respond to it promptly to minimise crop loss effectively.
Agri-input companies: Detecting the crops under cultivation allows agri-input companies to determine the regions or farms that would benefit the best from their inputs. Organisations that manufacture crop protection products can optimise their sales specific to the target crop and their growth stage, while farm machinery companies can improve farmer engagement by reaching out to them at the right states of cultivation.
Government agencies: Crop-cutting experiments are now made time- and cost-efficient with the use of crop detection and identification of crop stages at a regional level. Real-time insights improve visibility throughout the cultivation period and also enables government agencies to obtain fairly accurate estimates of crop yields to help officials plan food supply better and speed up insurance claims.
Insurance companies: Satellite image processing, coupled with deep learning, permits agri-insurance providers to assess crop losses due to natural calamities more accurately, helps overcome the many shortcomings of the manual procedures, and reduces the resources required by them for the entire process.
Lending institutions: SmartRisk’s ‘agri-worthiness report’ provides banks with a detailed summary of the crop over the last five seasons for a specific plot. The report makes it possible for the institutions to appraise loan requests and preemptively assess NPA based on the crop(s) cultivated by the farmer previously, the estimated yield, and relative growth index. Officials can also analyse crop growth in real-time by utilising this alternate agri-data.
NGOs and development agencies: Organisations that enable agriculture, especially in developing or underdeveloped countries, can leverage crop detection capabilities to map cultivation of crops across regions, monitor their health in real-time, and provide farmers with advisories to improve the productivity or prevent extensive damage to crops due to diseases, pest infestation, or unexpected weather conditions.
