ERIC: Estimating Rainfall with Commodity Doorbell Camera for Precision Residential Irrigation

TL;DR

ERIC addresses hyperlocal rainfall uncertainty in residential irrigation by estimating rainfall from commodity doorbell camera footage at the edge. The system combines lightweight visual/audio features with an edge ANN (and an optional cloud CNN) to detect and quantify rainfall and to drive irrigation scheduling via ET-based water balance. It is implemented on a Raspberry Pi 4 ($75) and deployed across five homes with 750 hours of video, achieving a rainfall estimation error of approximately $\sim 5\ \mathrm{mm/day}$ and significant water savings of $9{,}112$ gallons per month ($28.56/month$). The work demonstrates privacy-preserving, low-cost, automated precision irrigation, with open data and code.

Abstract

Current state-of-the-art residential irrigation systems, such as WaterMyYard, rely on rainfall data from nearby weather stations to adjust irrigation amounts. However, the accuracy of rainfall data is compromised by the limited spatial resolution of rain gauges and the significant variability of hyperlocal rainfall, leading to substantial water waste. To improve irrigation efficiency, we developed a cost-effective irrigation system, dubbed ERIC, which employs machine learning models to estimate rainfall from commodity doorbell camera footage and optimizes irrigation schedules without human intervention. Specifically, we: a) designed novel visual and audio features with lightweight neural network models to infer rainfall from the camera at the edge, preserving user privacy; b) built a complete end-to-end irrigation system on Raspberry Pi 4, costing only \$75. We deployed the system across five locations (collecting over 750 hours of video) with varying backgrounds and light conditions. Comprehensive evaluation validates that ERIC achieves state-of-the-art rainfall estimation performance ($\sim$ 5mm/day), saving 9,112 gallons/month of water, translating to \$28.56/month in utility savings. Data and code are available at https://github.com/LENSS/ERIC-BuildSys2024.git

Figures (18)

• Figure 1: Comparison of the current weather-based irrigation system (left) with our ERIC system (right). Current weather-based systems obtain rainfall data from nearby weather stations where the rainfall intensity can differ significantly from the hyperlocal rainfall at the residential sites, resulting in a significant waste of irrigation water. Instead, our ERIC system obtains accurate hyperlocal rainfall estimation from a doorbell camera, significantly improving irrigation precision.
• Figure 2: Our field experiment shows that rainfall measurements from a nearby weather station that is only 1.7 miles away can differ as much as 54% (43 mm on Aug 15) from the true hyperlocal rainfall as measured by a rain gauge at the residential site. However, our ERIC system estimates hyperlocal rainfall accurately, saving over $9,000 gallons of water for August 2021 (cf. Fig. \ref{['fig:water_saving']}).
• Figure 3: Left: an illustration of weather-based scheduling methods by considering the water balance between incoming water (rainfall, irrigation) and outgoing water (soil evaporation, plant transpiration). Right: a partial map of the industrial state-of-the-art weather-based program (WaterMyYard wmy) shows large spacing between weather stations, resulting in imprecise rainfall data for irrigation scheduling.
• Figure 4: ERIC system architecture. ERIC system harnesses the existing doorbell camera to stream the video to the irrigation controller board and then leverages a lightweight pretrained model on the board to infer rainfall intensity at the edge using both visual and audio features. Next, ERIC optimizes irrigation based on the estimated rainfall, $ET\_loss$ retrieved from nearby weather stations, and plant/soil types from user input on the smartphone App. Finally, it activates sprinklers according to the optimized schedule. ERIC also integrates Alexa for voice-controlled irrigation and allows users to upload videos to the cloud to use more powerful CNN models.
• Figure 5: Our rainfall estimation workflow at the edge. Our pipeline extracts visual and audio features from the input video and feeds concatenated features to a rain detector and estimator to predict raining minutes and rainfall intensity, which are then aggregated into cumulative daily rainfall.