Automatic Soil Sampling Site Selection in Management Zones Using a Multi-Objective Optimization Algorithm

Abstract

Precision agriculture hinges on accurate soil condition data obtained through soil testing across the field, which is a foundational step for subsequent processes. Soil testing is expensive, and reducing the number of samples is an important task. One viable approach is to divide the farm fields into homogenous management zones that require only one soil sample. As a result, these sample points must be the best representatives of the management zones and satisfy some other geospatial conditions, such as accessibility and being away from field borders and other test points. In this paper, we introduce an algorithmic method as a framework for automatically determining locations for test points using a constrained multi-objective optimization model. Our implementation of the proposed algorithmic framework is significantly faster than the conventional GIS process. While the conventional process typically takes several days with the involvement of GIS technicians, our framework takes only 14 s for a 200-hectare field to find optimal benchmark sites. To demonstrate our framework, we use time-varying Sentinel-2 satellite imagery to delineate management zones and a digital elevation model (DEM) to avoid steep regions. We define the objectives for a representative area of a management zone. Then, our algorithm optimizes the objectives using a scalarization method while avoiding constraints. We assess our method by testing it on five fields and showing that it generates optimal results. This method is fast, repeatable, and extendable.

A New Dissimilarity Metric for Anomaly Detection in Management Zones Delineation Constructed from Time-Varying Satellite Images

“…Then, the satellite data are downloaded from Sentinel-2. We used Sentinel-Hub API in this regard [10]. After that, the annual performance data are calculated, and zones are delineated automatically based on our zoning algorithm.…”

“…Moreover, our framework demonstrates versatility in its compatibility with various performance functions. While studies such as [10,37] utilized performance metrics like peak green or maximum pixel-wise NDVI to capture maximum productivity potential, these metrics may not fully encompass vegetation dynamics throughout the growing season. Additionally, relying solely on one satellite image exposes the analysis to vulnerability from environmental factors such as precipitation and cloud cover during peak growth times.…”

“…Additionally, they demonstrated that the NDVI becomes particularly valuable in cases where yield is not a clear discriminator, aiding in the determination of zone labels. In a complementary approach, Kazemi and Samavati [10] utilized historically normalized NDVI values to generate a fertility map for a field, subsequently using this map to partition the soil into distinct zones. However, challenges such as obtaining cloud-free images persist [25], leading to the exploration of methods like approximating the cloud-covered areas in the images [26] or data averaging [15] to mitigate these limitations.…”

“…In this section, we describe the overview of the chosen zoning technique. Our approach is akin to that in [10] and involved identifying zones based on historical performance. We constructed the performance indicator based on satellite imagery using an algorithm similar to that in [37].…”

“…In our experiments, we generated the performance indicator using historical NDVI data captured on the greenest day of each growing season and named them Peak Green. With one growing season annually, we obtained one peak green image yearly that was used as the growth indicator and included NDVI values at the peak day of each year [10]. To capture abnormal years, we introduced the Zoning Dissimilarity Metric (ZDM), which can capture the dissimilarity among peak greens.…”