A Decision Support Framework for National Crop Production Planning Based on Mosaicing, Historical Data and Satellite Imagery

Abstract

The thesis proposes a framework that optimizes the profit from crop production considering all the user’s constraints by taking into account field-level data and imagery from Unmanned Aerial Vehicles (UAVs) and satellites. The framework can serve a farmer and can also be scaled to work at national levels. The problem is divided into multiple levels. Image processing, machine learning, and integer linear programming-based models are used to address each level of the problem. While developing the solution we found out that most of the available videos are from legacy UAVs that have low quality. We also realized that these systems are still operational, they cannot be improved or retired owing to budget constraints. We decided to take up the challenge of using these videos to generate a mosaic of the agriculture area for further analysis. To create a mosaic of the area, our developed system first separates scenes, separates feeds from multiple on-board cameras, and stabilizes and enhances each feed to generate a super mosaic by merging all camera feed mosaics for a single scene. We find these mosaics very useful in defining field boundaries, identifying objects, and determining the ground-truth about adjoining land segments. The extracted information from mosaic processing is used to estimate the exact cultivated area that is subsequently used for yield prediction. Satellite imagery from Sentinel-II and LANDSAT-8 are used for this study to calculate different vegetation indices. The framework stores these indices to create a knowledge-base over the time period of crop growth. These indices are combined with historical meteorological data and crop-specific inputs to estimate crop yield using Machine Learning. This data of yield prediction for different crops in a specific area is used in the mathematical model that optimizes profit while satisfying a set of constraints. The solution of the model recommends a specific crop that shall be cultiiv vated in that area. National need and export demand are the key constraints set by the user. The proposed solution not only helps farmers and the government reap profits, but it can also help the logistics and supply chain managers to timely manage agricultural operations such as the availability of seeds for the recommended crops, fertilizer, pesticide, water, storage, and other aspects, as well as supplying mills with an estimate of the approaching harvest. The framework is validated using historical data from many sectors of Pakistan’s cultivated lands, with trials on single farmland with multiple fields, as well as trials conducted at multiple sites and fields in different geographical areas. The results demonstrate that the solution from the framework can help the government to optimize profit from the crops while meeting the national needs