Current farmland scouting methods, mainly including manual scouting and sporadic low-resolution imaging techniques, fail to provide near-real-time, unmanned and precisely-localized detection of biofuel crop stressors, such as herbivores and weeds attacks, ultimately causing significant losses in crop yields up to 50% and leading to overspray of chemicals everywhere. General statement of how this problem is being addressed: We plan to construct an initial numerical model and a high-speed and accurate ML-assisted computational prediction model of physical phenomenon (gas concentration spread and changes) and economic analysis under various environmental factors, (2) verify the developed algorithm with small-scale (<1 acre) controlled experiments (>100 repetitions) utilizing the previously-developed sensor network, and (3) develop a user interface app with a merge of an economical prediction model.What is to be done in Phase 1: During the phase 1, this STTR proposal aims to develop a computational model with a machine learning algorithm to enable the unmanned identification of near-real-time localization of weed/herbivore attacks in biofuel crops (sorghum) yield with minimal numbers of distributed sensors. Specifically, the aimed computational model will convolute several parameters through a machine learning model over a farming season for accurate prediction of the crop damage points including the location, timing and orders in hexanal detection within the sensor network, wind speed and direction, temperature, and humidity data in order to maximize the accuracy with minimal numbers of distributed sensors. It will further refine its interface with individual farmers by establishing user-friendly alerting and field status reporting. More importantly, the computation model integrated with the ultra-low power sensor network will also provide economical gain status monitoring by considering the status of reduced labors, farming supplies and equipment maintenance. Previously, our team developed and demonstrated innovative near-zero-power and thus long-term deployable gas (hexanal) sensors in completely off-grid sorghum fields in Nebraska, enabling the first continuous monitoring of gas emission from the field without the needs for battery replacements. The key enabling technology was to make the gas sensors normally-dormant (thus nearly zero-power) but continuously monitoring. It subsequently demonstrated unmanned and near-real-time detection of the released hexanal from stressed sorghums utilizing the wirelessly-tethered network of distributed gas sensors in the actual sorghum field. This preliminary validation was performed in multiple sorghum fields in Nebraska and Puerto Rico, previously sponsored by the ARPA-E OPEN 2018 project. Building on top of these previous demonstration, this proposal plans to seek to develop a computational algorithm that enables the accuracy in pin-pointing (localizing) the location of stressed sorghums for rapid treatment by farmers, and optimizes the number of deployed sensors for economically sound solutions for actual sorghum farming. We believe that this development of such a computational model and the utilization of the developed low power sensor network will provide a key step toward autonomous smart farming and the paradigm shift in scouting crops, ultimately leading to significant enhancement in the production of biofuels as well as in the reduction of pesticides and costs significantly.This integrated system will lead ultimately enabling localized treatment, instead of over-spraying chemicals everywhere, recovering the current crop loss of 10~30% total biofuel crop production. This will help reduce the overall crop production cost by 10~30%. It will help the farming cost reduced, the biofuel crops and farm soils be much healthier due to lower dose of chemicals, and thus higher biofuel production efficiencies. Additionally, recovering the sorghum production loss due to this development will contribute to the sequestration of greenhouse gases (GHG) and improvement of fuel efficiency.
