Application of artificial intelligence to variable rate technology in agriculture

Abstract

Variable Rate Application (VRA) plays a pivotal role in enhancing agricultural profitability by optimizing the use of resources and promoting consistent crop growth. This also helps mitigate the negative environmental impact of farming practices. However, the implementation of VRA is heavily reliant on data. An effective VRA prescription involves an agronomist’s in-depth knowledge of the soil and crop conditions within Homogenized Management Zones (HMZs). Certain soil attributes like electrical conductivity, elevation, and soil moisture are measured using proximal sensors installed on farm machinery. However, other soil properties like soil texture and Soil Organic Matter (SOM) measurements require soil sampling and laboratorybased testing. Similarly, crop and weed information is gathered via manual scouting. The collected SOM, soil texture, and crop information based on limited sampling may not be representative of whole field conditions resulting in low spatio-temporal resolution of information. Our research seeks to bridge these gaps by proposing costeffective and scalable solutions that improve spatio-temporal resolution. We suggest installing RGB sensors on farm machinery to monitor crop and weed growth, categorize soil texture, and estimate SOM. This high spatio-temporal information gathered is subsequently processed to investigate if improved HMZs can be identified. We develop crop and weed-specific semantic segmentation methods to detect, localize and quantify crops/weeds, yielding a mean Intersection Over Union (mIOU) up to 83%. These semantic segmentation models are customized to handle agricultural image data, minimizing memory usage and computational costs during training and inference. Through this adaptation, we observe a 6% performance improvement in crop and weed semantic segmentation. The efficiency of binary semantic segmentation models is further enhanced by up to 12% using ensemble learning methods. We recognize the strong correlation between soil properties and crop/weed densities and thus use this relationship to our advantage. We train machine learning models to predict crop and weed densities based on soil properties and satellite data. To accurately predict SOM and soil texture from RGB images, we employ a hybrid approach that combines deep learning and conventional image-processing techniques to overcome the challenges posed by uncontrolled field conditions data. Lastly, we explore the potential for identifying HMZs based on resultant high-resolution crop and soil information. Parts of this research are successfully commercialized under the product name ”SWATCAM”.