Deep Learning for Soil and Crop Segmentation from Remotely Sensed Data

Abstract

One of the most challenging problems in precision agriculture is to correctly identify and separate crops from the soil. Current precision farming algorithms based on artificially intelligent networks use multi-spectral or hyper-spectral data to derive radiometric indices that guide the operational management of agricultural complexes. Deep learning applications using these big data require sensitive filtering of raw data to effectively drive their hidden layer neural network architectures. Threshold techniques based on the normalized difference vegetation index (NDVI) or other similar metrics are generally used to simplify the development and training of deep learning neural networks. They have the advantage of being natural transformations of hyper-spectral or multi-spectral images that filter the data stream into a neural network, while reducing training requirements and increasing system classification performance. In this paper, to calculate a detailed crop/soil segmentation based on high resolution Digital Surface Model (DSM) data, we propose the redefinition of the radiometric index using a directional mathematical filter. To further refine the analysis, we feed this new radiometric index image of about 3500 × 4500 pixels into a relatively small Convolution Neural Network (CNN) designed for general image pattern recognition at 28 × 28 pixels to evaluate and resolve the vegetation correctly. We show that the result of applying a DSM filter to the NDVI radiometric index before feeding it into a Convolutional Neural Network can potentially improve crop separation hit rate by 65%.

A novel NIR-image segmentation method for the precise estimation of above-ground biomass in rice crops

“…The method achieved an average accuracy of 92%. Others have used machine learning methods for plot segmentation, and classification [23,24]. All of these methods require training stages that limit real-time functionality.…”

Predicting Canopy Nitrogen Content in Citrus-Trees Using Random Forest Algorithm Associated to Spectral Vegetation Indices from UAV-Imagery

“…Recently, machine learning algorithms have been used for different remote sensing applications [14][15][16][17][18]. Algorithms like artificial neural networks (ANN), support vector machine (SVM), decision trees (DT), random forests (RF), and others are powerful tools in assisting in UAV-based image analysis [19].…”

Advancing High-Throughput Phenotyping of Wheat in Early Selection Cycles

“…This could be more pronounced in plots with fewer rows. Recent studies have demonstrated the possibility to improve the segmentation of plant-soil pixels, e.g., using Support Vector Machine (SVM) classification or Convolutional Neural Networks [27,33,34]; (ii) aerial-based sensing has an advantage over ground-based sensing platforms in generating surface maps in real time and measuring plant parameters from a large number of plots at a time, typically associated with the time required to make ground-based measurements in large trials [12,13]; (iii) using high-resolution and low-altitude UAVs can overcome further limitations of ground-based sensing platforms, such as the non-simultaneous measurement of different plots, trafficability, row, and plot geometries requiring specific sensor configurations, and vibrations resulting from uneven field surfaces [12,28]. Given that the operation of UAV image acquisition is less labor-intensive, and owing to improved segmentation procedures and a higher precision than non-imaging proximal sensing, aerial-based multispectral sensing via UAV is expected to increase the efficiency of high-throughput phenotyping in large-scale plant breeding programs [10,12].…”