Accurate identification and segmentation of arable plots are of great significance in agricultural production. However, the traditional solution identifies plots through the spectral analysis of remote sensing images, which not only requires large bodies of sample data in the early stage and repeated collection of the same crop in different areas but also fails to present desirable accuracy.

Multispectral remote sensing image

To address this problem, SINO-ECO adopts PaddleSeg, a development kit for image segmentation in PaddlePaddle’s deep learning open-source framework, to deal with remote sensing data and extract farm land area, thus assisting relevant departments and practitioners in yield estimation.

Technical Solution

Case Design

• Optimal models for prediction accuracy

A combination of DeepLabv3+ and Xcep-tion65[zy1]  backbone networks is adopted considering the high prediction accuracy required by plot segmentation.

• Four-channel images for data enhancement

Paddle Seg [zy2] supports the training of four-channel spectral images to cover the shortage of normal RGB ones. Apart from the three RGB image channels, an additional NIR (Near Infrared Ray) is stitched to generate a four-channel image.

• Tuned training for accuracy improvements

The DeepLabV3+ model, by configuring the model to support four-channel images in the configuration file, is trained to learn the features of NIR spectral reflection to identify the vegetation, which can bring a 5% improvement to the final plot segmentation accuracy [zy3] of 90%.

The Network Structure of DeepLapv3+

Actual Effects

A series of optimization have been achieved with the algorithm in place.

•  The accuracy rate[zy4]  of agricultural parcel identification has increased to over 90%.
•  The highly adaptable model requires no data recollection even with changing areas, which greatly cuts labor input and simplifies plot information collection and analysis.
•  The intelligent extraction of arable land information from remote sensing images enables the real-time monitoring and analysis of agricultural remote sensing data.
•  A recognition accuracy exceeding 90% in the pest and disease analysis of crops.

Diagram of agricultural parcel identification

Company Profile

SINO-ECO, founded in 2014 and supported by the national 863 and 973 programs, firmly upholds the concept of "data creates value, and sci-tech for agriculture, rural areas, and rural residents". It has built a platform serving as an engine for agricultural digital map applications featuring full-dimension and high precision based on a model that integrates satellite, meteorology, ground spectrum, and crops for real-time CT monitoring, prediction, and decision making.