CoPEST

Being a major cash crop of Pakistan, cotton is considered the backbone of the economy. The cotton and textile sectors contribute almost half the country’s industrial base and cotton is the principal cash crop of Pakistan, providing critical income to the country’s households. Despite its significance to the economy, the cotton production has decreased over the last years. The low production of cotton has also drastically hampered the Pakistan textile industry thereby considerable loss in  GDP every year.
There are several causes for such low production. The insect and pest attacks are the major reason identified for the reduction of yield and the quality of the cotton bales. The cotton crop is severely affected by pest attacks. Out of 100% of the pesticides utilized in all the crops in Pakistan, 80% of pesticides are sprayed on the cotton crop. This  indiscriminate use of pesticides also affects the quality of cotton bales. Because of the requirement of fertilizers and plant protection from pests, the production cost of cotton also becomes high. Constant utilization of insecticides has necessitated the adoption of integrated pest management (IPM) approaches in cotton crops. A sustainable IPM
approach could lower the reliance of the farming community on pesticides.
In cotton, the pest  infestation caused deterioration in lint quality and 10–40% losses in crop production (Gahukar, 2006). Early prediction of pest attacks can be very helpful in improving productivity in agriculture. To this end, we propose to develop an AI-based IoT-enabled system for early detection of cotton pests using the data of environmental parameters  and multispectral images acquired from multiple sensors, which help to alert the farmer of taking prevention measures hence producing high crop yield. Several environmental  parameters have been highlighted in the literature that is significant for plant growth.

The data from environmental parameters, which include the air temperature, air humidity, CO 2 concentration, Illumination Intensity, Soil Moisture, Soil Temperature, Leaf  Wetness, and Soil Humidity, and the sensory drones will be stored on cloud-based services and the Deep learning models are developed to analyze time series-based sensor  fusion data for early prediction of cotton pest.
The deep learning framework is based on deep mutual learning (DML) strategy to learn from multi-sensor fusion data. The transformer-based architecture will be used for time  series sequence modeling to identify hidden patterns in environmental factors and learn their association with pest prevalence. The object detection convolutional neural  networks (RCNN or YOLO family) will be used to detect pests’ prevalence in multispectral drone imagery. This will enable end-to-end joint learning on multi-model embeddings  obtained from different network branches.

Additionally, the health of cotton crops on a large scale will be mapped which identifies the risk fields and regions for specific pests and a timely warning can be issued to the local farmers to prepare for such pest attacks. Finally, a dashboard shall be developed along with the web portal that provides analytics to the agronomist to find the correlation of multiple factors with cotton pest prognosis.