Feasibility of a low-cost weather sensor network for agricultural purposes : a preliminary assessment

Abstract: This study has focused on challenges encountered when setting up a weather-network for agricultural purposes (e.g. linking temperature to the suitability of crops and pest incidence) with only low-cost sensors and materials. The study included a set of experiments in a meteorological station and a one-month period of observations in a large coffee plantation with a complex terrain in Costa Rica. The created network is intended to be linked to agronomic trials, and are part of a package that will make farmers the scientists in a range of extension projects. Ongoing projects provide farmers with a range of seeds that are tested, while feedback is provided by a crowd-sourcing approach. Farmers send text messages back to the project managers. The coordinates of the farmers' field can be extracted from a stack of rasters with climate information, resulting in a clear understanding of the conditions during the agronomic trials. Experiments with different sensor (iButton DS1923) resolution showed that losses in precision when using a low temperature resolution are small compared to other losses (e.g. interpolation in time and space). Using a high-resolution for humidity observations provides very small improvements over low-resolution, as it provides data with the same accuracy. Experiments also focused on adjustment to PVC tubes, which functioned as sensor-shielding. All adjustments provided large differences with a certified shielding for the maximum temperature on sunny days. The best coating to limit the impact of radiation was insulating foil, and it is recommended that future experiments focus more on aeration, as this has not yet provided the expected benefits. As no (combination of) adjustments provided data in line with the reference station, different types of data calibration were tested. While direct correction of data by a polynomial regression model provided reasonable results, the main difference between PVC shields and the certified sensor was caused by the faster heating/cooling over time (thermal-inertia properties). By creating a linear model between change in time in the PVC and certified shield, a calibration model was developed that has been used to correct data. This has been done by setting an anchor point on each day, to which the corrected change was added/subtracted. With some minor additional calibration, this model provided data that was very similar to data in the certified shield. After the initial experiments were analysed, one hundred sensors were placed in a large coffee plantation with a 500 meter elevation gradient; fourteen sensors were lost and six provided incorrect data. The correlation between temperature and a range of variables was assessed. This included static (elevation, slope, aspect, canopy height, leaf-area-index and daily radiation) and dynamic (hourly radiation) covariates. On average, 52% of variance in temperature could be explained by static covariates. Including hourly radiation as covariate instead of daily radiation improved this model by 1%. Elevation is by far the most important independent variable (± 67%), although this influence is lower during periods with high temperature. A higher daily maximum temperature reduces the strength of elevation/temperature correlation. These are periods during which temperature is harder to predict based on interpolation in the complex terrain. A lower correlation between elevation and temperature can partly be compensated by a stronger correlation with other covariates; hourly radiation contributes on average 20% to the temperature-predicting models during hours with sun (although the models can only predict 54% of variance during these hours). Geostatistical interpolation has been tested for 80, 40, 20 and 8 sensors, with different kriging approaches and sets of covariates. Cross-validation provided the best results for universal kriging with elevation. Dynamic kriging provided smaller errors only with the full 80-sensor network. Co- and Spatio-Temporal kriging provided larger errors in predicting a left-out sensor, while data of the sensors included in the kriging showed least modification. The preferred approach depends on the network objective and reliability of data. While the network in this study cost ±US$ 8,700, a sufficiently accurate network of 25 sensors, can be created with a smaller budget: 20 low-res (temperature) iButton sensors (DS1922L-F5), 5 high-res (temperature and humidity) iButton sensors (DS1923-F5), 50m thin white PVC, 50 PVC elbows, 1m2 insulating foil, a small amount of fibre-glass mesh, and labor for construction (drilling holes and assembling). The cost for this weather network - which can store 341 days of 1-hour resolution data - will be approximately US$ 1,450. A 50-sensor network would still cost